Beyond Free Will: Understanding Approaches to Agency and their Suitability for Bandersnatch-like Titles

Abstract

When Bandersnatch (2018) was released on Netflix, interactive storytelling became accessible to a mainstream audience on a new scale. While this interactive film lets audiences make binary choices, the influence they have over the plot is limited, and at times the correlation between a choice and the resulting story is difficult to recognize. Although it can be argued that this constitutes a thematic design choice for this particular title, we think there is general room for improvement for this type of highly restrictive, branching structure film, in order to make the format applicable to a wider range of themes and stories. In this paper, Bandersnatch is examined as a representative of its format in order to develop and identify approaches to increasing agency. We use Hartmut Koenitz’s SPP model to understand the title and its format, and the hermeneutic strip extension to assess the perceived agency. Then, we introduce and examine potential approaches to increased agency from other interactive narratives to understand their adaptability and impact. Our discussion concludes that the most promising design idea for increased agency in Bandersnatch-like titles is invisible agency; an approach in which a player model is generated based on identifiable traits in the audience behaviour and used to select matching plotlines. This approach would allow audiences to see the results of their choices immediately, but also allow the impact of the choices to accumulate as the plot progresses, thereby increasing the overall sense of agency.

Full text

Entertainment Computing 43 (2022) 100500
Available online 20 May 2022
1875-9521/© 2022 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Beyond free will: Understanding approaches to agency and their suitability
for Bandersnatch-like titles
Anna Marie Rezk
a
,
b
,
*
, Mads Haahr
b
a
School of Design, Edinburgh College of Art, University of Edinburgh, Edinburgh, UK
b
School of Computer Science and Statistics, Trinity College, University of Dublin, Dublin, Ireland
ARTICLE INFO
Keywords:
Interactive digital narrative
IDN
Agency
Invisible agency
Obscured agency
Involuntary agency
SPP model
Double-hermeneutic-circle
Hermeneutics
Black Mirror Bandersnatch
Kaleidoscopic form
Replayability
Player model
Narrative momentum
ABSTRACT
When Bandersnatch (2018) was released on Netix, interactive storytelling became accessible to a mainstream
audience on a new scale. While this interactive lm lets audiences make binary choices, the inuence they have
over the plot is limited, and at times the correlation between a choice and the resulting story is difcult to
recognize. Although it can be argued that this constitutes a thematic design choice for this particular title, we
think there is general room for improvement for this type of highly restrictive, branching structure lm, in order
to make the format applicable to a wider range of themes and stories. In this paper, Bandersnatch is examined as a
representative of its format in order to develop and identify approaches to increasing agency. We use Hartmut
Koenitz’s SPP model to understand the title and its format, and the hermeneutic strip extension to assess the
perceived agency. Then, we introduce and examine potential approaches to increased agency from other
interactive narratives to understand their adaptability and impact. Our discussion concludes that the most
promising design idea for increased agency in Bandersnatch-like titles is invisible agency; an approach in which a
player model is generated based on identiable traits in the audience behaviour and used to select matching
plotlines. This approach would allow audiences to see the results of their choices immediately, but also allow the
impact of the choices to accumulate as the plot progresses, thereby increasing the overall sense of agency.
1. Introduction
After the launch of Black Mirror’s Bandersnatch on the online
streaming platform Netix in December 2018, the format of interactive
lms has become accessible to a mainstream audience. After the title’s
success, Netix announced its plans to produce more interactive titles
aimed at a mature audience [1], such as Unbreakable Kimmy Schmidt:
Kimmy vs. the Revered (2020) [2]. This format is also being explored by
smaller production companies, like CrtlMovie who are dedicated to
creating interactive lms with experimental and sophisticated UI, like
Late Shift (2016) [3].
While Bandersnatch has brought interactive lms to a wide audience,
its format is subject to severe limitation in terms of audience agency.
Therefore, a discussion on how the format could potentially be improved
is essential, timely and constructive for the production of interactive
works in the future.
This paper takes its starting point in our previous work which
appeared in the ICIDS 2020 proceedings [4] and aims to identify design
ideas that can increase audience agency in Bandersnatch-like titles
without jeopardizing the experience in terms of its narrative mo-
mentum, replayability, and thematic elements. Due to its prominence
and typicality, Bandersnatch is chosen as a representative of its format of
highly restrictive interactive lm in the case study of this paper. Our
methodology is to examine Bandersnatch with the SPP model [5] and its
extension [5,6] in order to understand the title’s narrative format and
structure. This will be followed by an analysis of the title’s user agency
or (as audiences might perceive it) the “free will” it gives to its players in
their choices. However, these two concepts will not be used synony-
mously in this paper, because by using them interchangeably in the
realm of interactive narrative, one ignores the importance of necessary
constraints imposed on the audience to drive the story forward [7].
Therefore, it must be considered that increased agency through
increased interaction could potentially lead to a decrease in the narra-
tive momentum. Once the audience becomes in charge of the story, the
Abbreviations: IDN, Interactive Digital Narrative; SPP, System, Process, Product; UI, User Interface; ICIDS, International Conference on Interative Digital Sto-
rytelling; CYOA, Choose Your Own Adventure.
* Corresponding author.
E-mail address: a.rezk@ed.ac.uk (A.M. Rezk).
Contents lists available at ScienceDirect
Entertainment Computing
journal homepage: www.elsevier.com/locate/entcom
https://doi.org/10.1016/j.entcom.2022.100500
Received 15 February 2021; Received in revised form 20 December 2021; Accepted 15 May 2022

Entertainment Computing 43 (2022) 100500
2
risk increases of them getting stuck at certain plot points, thereby pre-
venting the story from unfolding or developing in any meaningful way.
These potential risks have to be taken into consideration when sug-
gesting ways to increase agency in interactive storytelling.
After the analysis of the title, different approaches to agency will be
outlined and assessed regarding their suitability for titles with a similar
format to Bandersnatch, and similar target audience and screening set-
tings. Finally, the most suitable approach will be elaborated, and its
conceptual implementation will be explored further.
2. Background
2.1. Top-down and bottom-up systems
Marie-Laure Ryan differentiates between two fundamental ap-
proaches in interactive narrative: The bottom-up, emergent systems that
create stories on the y, and the top-down systems that rely on pre-
scripted content. Examples of bottom-up systems are simulation games
like The Sims, where, as Ryan explains,
the player’s selection counts as the performance in the ctional
world of the action described by words on the menu […] The suc-
cession of choices writes the life story of the Sims family [8, p. 50].
Here, the system has to react to the player’s behaviour in real time
and offer meaningful consequences to them. A more sophisticated, and
purely hypothetical, example of an emergent system would be the
ctional Holodeck machine from the television franchise Star Trek,
which was proposed by Janet Murray in 1997 and since then has served
as a guiding metaphor for researchers in interactive storytelling. The
Holodeck is a stage that allows users to engage with virtual environments
and where every single input affects the environment and thereby the
narrative. Since it would not be possible to store all the storylines
created by the user’s input in advance, the only way for such a complex
system to work would be to compute the effects in real time [9].
In contrast, top-down systems like Bandersnatch require all the scenes
that can be unlocked by the user to have been produced in advance. This
means that while the audience is given some agency over the narration,
they cannot truly create different endings but only unlock what is
already there. The resulting difference between these two approaches is
that while the emergent system can be run multiple times, creating
multiple outcomes, the top-down approach does not renew itself, even if
it offers some different narratives [8].
A drawback pointed out by Ryan concerning the bottom-up
approach, however, is the potential lack of closure, as she argues that
without the authorial control from the top-down approach, it is impos-
sible to create the Aristotelian curve of rise and fall of tension, or even
just a resolution and end of events [8]. An interactive drama to counter
Ryan’s concerns about the bottom-up approach is Façade (2005) by
Michael Mateas and Andrew Stern. It was considered a breakthrough
piece, as it interprets the player’s input in real-time to construct a
captivating dramatic arc while also allowing a high degree of interaction
with the AI-driven characters [10].
It is important to note that the top-down and bottom-up approaches
are not mutually exclusive and should be treated as two ends of a
spectrum, which allows elements from both ends to be combined in a
single experience. Sometimes in The Sims, the game takes control and
adds pre-scripted scenarios to the simulation, e.g., alien abductions.
While the game cannot control the state that the user has put the virtual
world in, such plot twists add narrative momentum to the simulation.
We consider this an attempt to drive the game and the story it tells
forward, even if simulation games like The Sims do not necessarily
require an overarching story.
However, this design opens up the opportunity to have emergent
systems exist within top-down systems, by allowing user behaviour to
generate events that would be consistent with top-down arcs. Obviously,
this would require limitations in the bottom-up interaction responses of
the system in order to be able to align the generated story world with a
pre-scripted top-down arc. Player modelling, i.e., detecting and under-
standing the player’s “cognitive, affective, and behavioural patterns”
[11,p. 45] during the interaction in playable scenes can be used to
determine the respective pre-scripted arc. A practical example would be
only adding the randomly occurring alien abduction sequence in The -
Sims, when the generated state of the universe and/or the player
modelling allow for it in a narratively meaningful way.
2.2. Interactivity and agency
One essential element of interactive stories is audience agency.
Murray describes agency as “the satisfying power to take meaningful
action and see the results of our decisions and choices.” [9,p. 159] If
players of a tabletop board game are given increased means of in-
teractions, like throwing dice and spinning dials, they might get the
sense of having an inuence on the experience. However, their actions
are neither chosen by them, nor do their effects mirror the player’s in-
tentions. This is where Murray draws the line between activity and
agency. As an example, she explains that a game of chess has a high
degree of agency even though it only offers few actions. That is because
all actions are “highly autonomous, selected from a large range of
possible choices, and wholly determine the course of the game” [9,
p. 161]. Murray denes agency in her glossary as follows:
When the behavior of the computer is coherent and the results of
participation are clear and well motivated, the interactor experiences
the pleasure of agency, of making something happen in a dynami-
cally responsive world. [12]
In order to allow for agency in a narrative, the narrative requires not
only multiple paths, but also oftentimes multiple endings. Depending on
how complex the story is meant to be, these formats often may not resort
to the win/lose simplicity of classic video games, but rather have mul-
tiple ending scenarios that can be understood as the consequence of the
player’s input. These outcomes can either be the direct effects of certain
actions or can be reached through a chain of uncontrolled scenarios
kicked off by the player’s input (similar to the buttery effect, a meta-
phorical example of how a tornado can be inuenced by something as
minor as the apping of a buttery’s wings months earlier).
There are many different structures for branching narratives, each of
which branches out differently and in different degrees, thereby allow-
ing different degrees of agency. Sam Kabo Ashwell has created an
extensive list of possible structures, some of which are the following:
The Time Cave structure (see Fig. 1) is the most obvious kind of
branching structure, where each decision point offers a new forked
pathway, thereby having the plot branch out exponentially. This struc-
ture strongly encourages replay, as different walkthroughs tend to be
substantially different in content and overall experience [13].
A less production-heavy alternative is the Gauntlet structure (see
Fig. 2). This structure is dened by its relatively linear thread that has
several branches which lead either to dead ends, backtracking, or a re-
joining with the central thread. Overall, this structure tells one main
story, which can either be enhanced with optional content or cease
prematurely if a dead end is chosen.
The Branch and Bottleneck structure (see Fig. 3) branches out at times
and comes back together for key plot points. Agency is facilitated in this
structure by the implementation of state-tracking, meaning that even
Fig. 1. Time Cave Structure [13].
A.M. Rezk and M. Haahr

Entertainment Computing 43 (2022) 100500
3
though the player ends up at a certain plot point, regardless of the
previous choices, these previous choices and the underlying behavioural
pattern of the player are stored in the system and later inuence the way
the story moves forward (i.e., only certain endings become a possibility
once a certain path was taken). Here, oftentimes players do not notice
the agency – in this case the agency is somewhat invisible.
The Loop and Grow structure (see Fig. 4) has one central thread that
keeps looping to the same point. Due to state-tracking each loop might
be slightly different than the previous with new options appearing and
others disappearing. This type of structure usually requires a thematic
justication for plotlines to be revisited over and over again, e.g., time-
travel.
3. Methodology
For this paper, Hartmut Koenitz’s analytical framework, the SPP
model and its extension (see Figs. 5–7), will be used to analyse Ban-
dersnatch. This analysis is essential to understand the IDN in question in
order to subsequently identify promising design ideas for increased
agency.
In Fig. 5, the term system is used to describe the interactive program
itself, including both the software and hardware required for the inter-
active experience. The process is the user’s interaction with the system,
which ultimately results in a product, a singular storyline based on the
user’s input, which would be different if the user’s input were to change.
The product is therefore an instantiated narrative [5].
As part of the model, Koenitz introduces three additional terms:
Protostory, narrative design, and narrative vectors. Protostory is the space
of potential narratives, “containing the necessary ingredients for any
given walkthrough.” It stands for both the code and the interactive
interface of the system, and thereby captures the
artistic intent that enables a participatory process of instantiation
resulting in the realisation of potential narratives […] The term
narrative design describes the structure within a protostory that
describes a exible presentation of a narrative [5,p. 99].
In other words, the narrative design deals with the sequencing of
elements and their connection in the narrations. A substructure of
narrative design are the narrative vectors, which provide specic di-
rections for the story. They have to be understood as substructures that
work in connection with the preceding and following parts of any
narrative. Their purpose, as Koenitz states, is “to convey important as-
pects to the interactor, to prevent an interactor from getting lost and to
aid authors in retaining a level of control” [5,p. 100]. A narrative vector
could, for instance, be a sudden event in the plot that shapes the
development of the story and can be compared to plot points in linear
narratives.
One thing that is not covered by Koenitz’s SPP model is the idea of
agency. However, in a paper by Christian Roth, Tom van Nuenen and
Koenitz himself, an extension to the model was introduced, namely the
hermeneutic strip or double-hermeneutic circle [6]. This strip aims to
illustrate the player’s narrative meaning-making process. It captures
both the interpretation of the system overall (i.e., the players’ reection
on what the system may allow and which freedoms or agency they have)
Fig. 2. Gauntlet Structure [13].
Fig. 3. Branch and Bottleneck Structure [13].
Fig. 4. Loop and Grow Structure [13].
Fig. 5. Koenitz’s SPP model [5].
Fig. 6. The SPP model’s associated terminology for the analysis of IDN [5].
Fig. 7. The additional extension by the double-hermeneutic circle offers a
methodological toolkit for the analysis of experienced agency [6].
A.M. Rezk and M. Haahr

Entertainment Computing 43 (2022) 100500
4
and the players’ interpretations of already instantiated narratives.
It is important to understand that this extension to the model bears in
mind that a player’s behaviour (which is assessed through their inter-
action with a system) is shaped by previous experiences made in the
interactive narrative. In short, past and present events inuence a
player’s future behaviour. As this extension will help evaluate agency
from a player’s perspective at predened points of the story, these key
points must be identied rst [6].
By examining the protostory, all assets of the title, as well as all
components that make up any storyline and the interface will be laid
out. Going into more detail, the narrative design will dene the seg-
mentation of different scenes and the choices that connect them. By
thoroughly analysing the narrative vectors, the level of control of the
producers will be understood, as these plot points are the orientation
points that prevent the audience/players from getting lost, resulting in a
loss of narrative momentum. The double-hermeneutic circle model will
help evaluate the agency from a user’s perspective at predened key
scenes. It is important to add that the authors of this extension used it in
the context of a quantitative analysis of players’ reactions based on Let’s
Play videos found online, where they examined reactions at a predened
key scene. In this paper, we base our analysis on the range of possible
reactions rather than a study of actual reactions.
4. Bandersnatch
Black Mirror’s Bandersnatch was marketed as the rst interactive lm
aimed at a mature audience on the streaming platform Netix. Similar to
previous episodes of the Black Mirror franchise, an introspective view on
technology is reected in the plot of the lm but also – for the rst time –
in its interactive structure. The critically scrutinized technology in
Bandersnatch is not only the one used by characters in the plot, but also
the one used by the audience itself, as the theme of the lm deals with
the question of the existence of free will, agency and control. The
controllable character, Stefan, an aspiring game developer in the 1980s
is attempting to create a video game called “Bandersnatch,” which has a
branching narrative structure, like the CYOA book it is based on. As the
lm progresses, the audience is forced to make choices that are not in
Stefan’s best interest due to the lack of more favourable options. In
response, Stefan nally breaks the fourth wall and confronts the audi-
ence by asking which outside force is controlling him. According to Roth
and Koenitz, in Bandersnatch, there is a parallelism for control, as agency
is “explored in parallel – in the diegetic world and the interactive
narrative experience” [14,p. 249].
4.1. Bandersnatch: system, process and product
In Bandersnatch, the audience can take control of the direction of the
plot. At predened choice points they are prompted to choose between
two binary textual choices within ten seconds. If no choice is made, the
system defaults to one of them automatically, allowing for what Roth
and Koenitz refer to as “passive consumption” [14,p. 249]. The overall
branching structure of the title is close to a Gauntlet, as introduced in
Section 2.2, since there is a main thread, which is close to linear, with
some branches emerging from it that lead to dead ends, as well as
backtracking and re-joining branches.
Using Koenitz’s SPP model for the analysis, the system in this title
consists of any device that the lm can run on, for the hardware part.
Netix has made some restrictions concerning which devices are
compatible with the lm and created an apology clip which is screened
in case the used device was incompatible with the functionalities of the
title. The software of Bandersnatch includes the interface, which is
shortly introduced and explained to the audience at the beginning of the
interactive experience. The UI at choice points consists of two textual
options and a timer in the form of a horizontal line which decreases in
length and disappears after ten seconds. Since the overall length of the
lm is subject to the audience’s input, there is no overall info of the
lm’s duration or a progress bar. Other parts of the system are all virtual
assets, which include all scenes, choice points and their respective op-
tions, as well as the program that manages the audiences’ inputs and
outputs the corresponding narratives. Another important feature is that
the system maintains state, as previous choices can be inuential or at
least referenced as the lm progresses.
The process in this title is determined by the audience’s input as well
as the choice options that the system provides. Due to the Gauntlet
structure of the lm, which contains multiple dead ends, the audience is
often prompted to change their chosen narrative path in something
equivalent to a respawn with Stefan saying, “I should try again.” Even
when a more conclusive ending has been reached, the system asks the
users if they want to explore yet another path. This option of going back
and trying out variations of the interactive narrative has been dened by
Murray as the “kaleidoscopic form,” which she describes as
the potential of interactive digital narratives to present us with
multiform scenarios in which the same events can be understood in
multiple contexts and the same starting points can be imagined as
giving rise to multiple possible outcomes. [15,p. 3]
Kaleidoscopic design goes hand in hand with the long-standing
notion of IDNs as replay stories [16]. Consequently, the longer one
chooses to interact with the lm or the more times one chooses to replay
it, the more storylines can be unlocked and the more likely one is to have
similar viewing experiences as other audiences. However, this in turn
means that the instantiated product is not as unique as Koenitz has
described it, by stating that “very different narrative products can
originate from the same system” [5,p. 98]. The structure of Bandersnatch
as well as the encouragement to explore as many narrative paths as
possible, does not allow for a high level of uniqueness or variety
regarding the instantiated product. The only major difference between
independent viewings would be the sequence in which the different
audiences have seen the different plot lines and endings.
4.2. Bandersnatch: protostory, narrative design and narrative vectors
The content in Bandersnatch can be broken down to the small entity
of scenes for the protostory. Another part of the protostory is the already
introduced interface and the code that allows the interactive lm to run
according to the audience’s input. Additionally, features of the default
interface of Netix are also available or adjusted to the nature of the
lm, such as the buttons to fast-forward and rewind ten seconds, which
only allow users to fast-forward until the next decision point but not
beyond it.
The narrative design of Bandersnatch is overall close to linear, which
is typical for the lm’s Gauntlet structure. It makes sense to segment the
design into bundles of scenes which are co-dependent. This means that
exibility exists between different bundles of scenes rather than within
them.
We can consider the narrative vectors to be the scenes that are
revisited upon respawning (i.e., letting Stefan “try again”) as well as the
respawn function itself, since they convey the message that the previ-
ously chosen path led to a dead end. Therefore, the narrative vectors act
as orientation points to guide the audience through the narrative. A
minimal impact on the narrative momentum is ensured through dra-
matic compression [17] upon respawning. This means that scenes that
have to be replayed when the audience decides to go back and try again,
are sped up until the decisive choice point is reached once more [18]. It
can be argued that the narrative vectors in Bandersnatch work in a way
that facilitates the consumption of the majority of storylines that exist.
4.3. Bandersnatch: double-hermeneutic circle and agency
To be able to use the double-hermeneutic circle to assess agency, we
must rst identify key choice points for the analysis [6]. It is important
to identify at least one choice point that any user would encounter,
A.M. Rezk and M. Haahr

Entertainment Computing 43 (2022) 100500
5
regardless of the narrative path they take. The logical choice falls on the
rst inuential choice point, which is when Stefan is offered a job as a
game developer. The previous two choice points were inconsequential
and even presented as such, e.g., choosing which cereal Stefan has for
breakfast. Therefore, the job offer, which Stefan can either accept or
decline, seems like it could result in two contrasting storylines.
This idea of two distinct branches resulting from this choice can
emerge from the upper hermeneutic circle’s interpretation of the system.
However, by analysing the bottom hermeneutic circle, one might
consider how previous choices were uninuential and therefore doubt
any signicant consequences to be brought about by this choice. Albeit
this choice point is presented as potentially plot-altering, it might
confuse the audience, given the previous choice points. In a meta-
discussion, it can be suggested that a novice IDN audience might
expect two major branching storylines from here onwards, whereas a
seasoned audience could be more aware of potential production con-
straints resulting in one of the options being a false choice. In this case,
the sceptical audience is correct, as accepting the job leads to an im-
mediate dead end and lets the audience redo this decision. This looping
behaviour then alters the hermeneutic circles, as the audience learns
that the only real choice is to decline the job offer.
One more key choice point worth examining is when Stefan is con-
fronted with two rather similar, destructive options, as briey intro-
duced in section 4. The audience is constrained to choosing whether
Stefan destroys his computer or pours his tea over it, when both options
clearly bear the same consequence of his work being lost. The lack of an
option that could lead to an alternative consequence lets the audience
realize that in this instance they have no real control over what is to
happen next, even if they were to inuence how it is brought about. With
tied hands, they can either surrender to the limitations of the available
options or watch the system default to one of the options for them. Either
way, it becomes a violation of the narrative contract through the
removal of agency that the audience expected to have up to this point.
However, in a cleverly constructed twist, Stefan breaks the fourth
wall by refusing his orders as he realizes that he does not wish to destroy
his computer and that he is being controlled by an outside force. Neither
a reection in the upper nor bottom hermeneutic circle could have
raised a suspicion of this plot twist, since all previous inputs by the
audience resulted in corresponding actions performed by Stefan. How-
ever, in one of the lm’s endings, an almost identical choice point ap-
pears, this time showing a grown-up Pearl after a time jump of roughly
30 years. As she is attempting to program Netix’s interactive lm
“Bandersnatch,” the lm that the audience is watching, she keeps
encountering errors and gets frustrated. Again, the presented options are
to destroy the computer or to throw tea over it. Due to the already
instantiated narrative, a reection in the bottom hermeneutic circle
might result in thinking that also in this instance will the given com-
mand be ignored by the system, especially since the audience has no
means of knowing that with this scene, they have reached an ending and
that it would therefore be unlikely to have another plot twist. However,
in this case, Pearl performs the destructive act respective to the selected
option.
Reecting on the overall agency in Bandersnatch, we observe that in
almost all choice points two options are given. One of the exceptions is a
ashback scene of Stefan’s deceased mother who asks him if he wants to
come with her on a journey. The only option is to say “No,” nodding at
the idea that the past is immutable. There are also a number of deceptive
choice points that offer two very similar options, like the two previously
discussed in which the audience can choose how Stefan and Pearl should
destroy their computers. Furthermore, the gravity of the choice points is
not held to a constant level, as they range from choosing which music to
enjoy to deciding whether or not to kill Stefan’s father. In the scene
where Stefan attempts to crack his father’s safe, each password option
the audience can choose from takes Stefan to a storyline that resonates
with the password. It is never revealed which password would have
opened the safe correctly; instead, the audience gets to unlock different
facets of Stefan’s life and the lm’s universe [19]. It is therefore not
clear, if Murray’s denition of agency being “the power to take mean-
ingful action and see the results of our own choices” [9,p. 159] works
well with Bandersnatch, as the audience’s input prompts a corresponding
output, but leaves the audience wondering about how meaningful the
consequence was. This can be due to the unrelatable personality of
Stefan, who the audience might not have had enough time to sympathise
with before having to control his life, or perhaps due to choice options
that violate the narrative contract, or perhaps even due to the unpre-
dictability of the Bandersnatch universe.
This unpredictability is seen in the choices that unleash a buttery
effect that causes several events, which the audience has no control over.
An example of this occurs early on in the lm when Stefan chooses to
accept the job offer, and the lm fast-forwards several months to reveal
the resulting game’s poor reviews. This example is symptomatic of the
overall inconsistency of the gravity of different choices, which is most
likely a stylistic choice to reect the lm’s theme of chaos and the lack of
control over life. However, this in turn means that the consequences of
the audience’s input are not always meaningful or foreseeable. This
conclusion is supported by the quantitative survey on Bandersnatch
conducted by Lobke Kolhoff and Frank Nack with a sample group of 169
participants who had seen the lm: Overall 69% have said they disagree,
or neither agree nor disagree with the claim that consequences in Ban-
dersnatch are foreseeable [20,p. 82]. Table 1.
5. Reection on replayability, narrative momentum, agency,
and thematic suitability
After having analysed Bandersnatch with Koenitz’s SPP model and its
extension, this section will evaluate its replayability, narrative mo-
mentum, as well as agency.
As established, Bandersnatch is structured around a main narrative
thread. The interface allows viewers to jump back to their previous
choices and alter them, and even once an ofcial ending has been
reached, the lm prompts its audience to decide if they want to go back
and explore an alternative storyline that had not been unlocked yet. In
this way, we could say Bandersnatch’s kaleidoscopic design trades the
overall per-title replayability for per-scene replayability, as it seems to
encourage the viewer to watch as many scenes as possible in one session.
In Bandersnatch, the narrative momentum is not inuenced by the
player’s input, as choices have to be made within ten seconds during
which the controllable character is shown to be reluctant about what to
do next. If the player does not make a decision within the allotted time,
the lm defaults to one of the options.
When taking a deeper look at Bandersnatch, the interaction is facil-
itated by two textual prompts that lead to different actions of the pro-
tagonist and thereby change the plot. However, as previously discussed,
there is a decision point at which Stefan refuses to follow through what
he is being ordered to do by the audience and breaks the fourth wall by
confronting the force that he thinks is controlling him – an act that ac-
cording to Roth and Koenitz also breaks the viewer’s identication with
Stefan and encourages a reection on agency [14]. In this twist of
events, the plot develops in a direction that is independent of the user’s
input, and what follows are streams of unexpected events, one of which
Table 1
Kolhoff and Nack’s questionnaire results on agency in Bandersnatch (SA =
strongly agree; A =agree; NA =neither agree nor disagree; D =disagree; SD =
strongly disagree) [20].
SA A NA D SD
Content frequency choices 20% 47% 20% 11% 4%
In control experience 21% 33% 24% 15% 7%
Desirable consequences 5% 26% 46% 20% 4%
Foreseeable consequences 5% 14% 33% 36% 11%
No choice while preferred 14% 30% 27% 20% 9%
Unnecessary choices 17% 24% 22% 28% 8%
A.M. Rezk and M. Haahr

Entertainment Computing 43 (2022) 100500
6
involves Stefan killing his father. This is a signicant act that is out of the
audience’s control in some storylines, while in other playthroughs the
player gets to choose to either kill the father or back off during the same
scene. Close to one of the endings, the audience can see Stefan proudly
talking about his video game to his psychologist Dr Haynes and saying
that he had nally nished it by reducing the amount of agency given to
the player: “Now they only have the illusion of free will, but really I
decide the ending.” This quote by Stefan seems to go hand in hand with
the structure of the interactive lm, as multiple different paths can lead
the audience to the same ending and some paths seem rather forced, like
the one where Stefan unexpectedly kills his father.
Moreover, the impact of the audience’s inputs varies signicantly in
Bandersnatch, as some choices can be entirely ignored by the system and
manoeuvred around to have the same output as the option that was not
chosen, such as when Stefan decides against taking LSD, but his tea is
spiked anyway. Other decision points offer two very similar options that
would result in the same consequence but bring it about differently,
which goes against the notion of agency and thereby further limits the
perception of it. As Sercan S
¸engün explains, “Forcing a choice and
constraining the alternatives or presenting inconsistent alternatives may
thwart instead of support the feeling of freedom” [21,p. 184]. Some
decision points let the player make rather trivial decisions while others
can become a matter of life and death. However, the trivial decisions can
bring about unforeseen consequences, either in a buttery effect or by
immersing the audience into the parallel realities that this interactive
lm is trying to fabricate.
This theme could arguably be the work reecting on the chaotic
nature of the universe, highlighting that much of the experience of
control in life is at best precarious and subject to a highly unpredictable
universe or at worst completely illusory. This thematic-structural reso-
nance works well, as the overarching question raised throughout the
narrative is whether the character is in control or not. The resonance is
also reected in the ctional CYOA book “Bandersnatch” on which
Stefan bases his video game with the same title. Here the enemy, the
“Pax” demon, is the thief of destiny who appears to be in control over the
plot. After having nished his book, the ctional author Jerome F. Davis
started believing he was being controlled by an outside force and that his
wife was spiking his drinks, making him more susceptible to being
controlled at “the behest of a demon called Pax.” This paranoia even-
tually led him to murder his wife, a tragedy that is discussed in one of the
rst scenes in the lm [22]. As the plot progresses, Stefan too starts
developing a similar paranoia, leading to the crucial scene where he asks
which force is in control, after refusing to follow the input of the
audience.
The similarity in the growing paranoia of Jerome F. Davis and Stefan
is of course striking, and it is worth noting that their respective “de-
mons” Pax and PAC (the abbreviation for the lm’s “Program and
Control” experiment) are also near homophones. In one plotline, Stefan
discovers that he is a subject of the PAC experiment, meaning that his
entire reality was fabricated by scientists and actors, even the trauma-
tising milestones, such as his mother’s death in an accident. In a
Truman-Show-like twist, Stefan learns that everything in his life was
controlled by the unethical research scheme that has distorted his
perception of reality and control throughout his life.
The theme is also hinted at visually, with the symbolism of the
branching narrative structure following the protagonist throughout.
With the existential questions about free will and control becoming the
focal point of the lm, it appears appropriate to have a branching
structure that undermines the audience’s sense of being in control of the
narrative at certain stages. Therefore, it can be argued that the thematic
choice of Bandersnatch is making a virtue out of its limited agency, by
justifying the inconsistencies in user agency with the theme, which re-
ects precisely on these concepts. However, this design for agency will
not generalize to works that try to engage with different themes. For this
reason, it is important to look at improving agency, so that the format
can express stories beyond those that have to do with free will.
6. Reection on different approaches to agency
In the past, agency has been given to audiences and users of IDNs in
different forms and to different extents. For instance, in Bandersnatch,
we may argue that agency is somewhat limited due to the binary nature
of the choices given in the top-down system, which in turn raises the
question about possible alternative approaches to agency. In Bander-
snatch, the extent of audience agency is quite obvious, as the interface
informs when a choice point is reached and gives two options to choose
from. The respective action takes place immediately after, making the
audience aware of the control they have over the lm.
By examining other IDNs, we can observe that approaches to agency
can have various forms and that agency can be granted through different
means. A different approach to agency is observed in the video game
Silent Hill 2 (2001), where the players have so-called “invisible agency,”
a term coined by Sercan S¸ engün. In this title, the system attempts to
assess and model the players’ psychological states based on their ten-
dencies and behaviour while playing by maintaining state and then ul-
timately unlocks one of the different endings accordingly [21]. While
the bulk of the game is largely linear in its narrative progression, the
endings are provided in in the same way as the ending in a typical
Branch and Bottleneck structure (see Fig. 3). In the case of Silent Hill 2,
the chosen branch is not an explicit choice but depends on cumulative
effects of a type of psychological player modelling performed during the
linear narrative part of the game. In short, as S
¸engün explains, “the
choices the player makes are actually projected tendencies and they
accumulate results in the long run” [21, pp. 183–184].
In the Mass Effect franchise (2007–2017), a similar approach to
agency is employed, as the decisions of doing side quests, behavioural
patterns as well as the engagement in dialogue trees inuence the
narrative. In the original Mass Effect trilogy, the chosen option from the
dialogue wheel would help assess the player’s morality and place them
on the path of a “Paragon” or “Renegade” accordingly. These paths have
immediate effects but also later alter the way the narrative progresses,
for example by affecting the available choice of allies as the game moves
forward [23]. The difference to Silent Hill 2’s invisible agency is that
Mass Effect’s dialogue wheel presents an obvious interactive interface,
and its agency therefore no longer seems invisible in all instances where
it is given, even if it also has consequences that are not immediately
obvious.
Another interesting IDN is the Danish interactive lm Switching
(2003). This lm is presented as a DVD and revolves around a strained
relationship between Frida and Simon, in which both appear to be stuck.
This title is unusual in that it has no on-screen interface to enable
interaction, but instead offers cues embedded in the dramatic perfor-
mance. To affect the story, viewers can press the Enter/OK button on
their DVD remote control when they wish to change the course of the
lm at choice points. These choice points can be recognized through
clues in the actors’ expressions, as they appear to dissociate or zone out
for a brief moment to allow the audience to take control of the plot. The
lm itself never comes to an end, as the audience gets stuck in a maze of
never-ending loops (see Fig. 4) of the relationship with the couple itself
[24]. The way that Switching is laid out allows for the choice between
passive consumption and active intervention at each subliminally pre-
sented choice point. However, users do not know what they would
achieve by intervening, nor do they know if they inuence anything,
since this format does not inherently offer feedback once an interaction
has taken place. Therefore, it can be argued that audiences only become
aware of their inuence after several replays during which they inter-
acted differently and observed narrative changes as a result. Conse-
quently, this approach to agency may be considered “obscured agency.”
For the enactive cinema installation Obsession (2005), audiences
were given a form of paradoxical agency, as they have no real control
over their choice but are aware that they have an inuence over the
narrative. With the help of biosensors, this interactive installation
measures physical responses of the audiences’ bodies to determine what
A.M. Rezk and M. Haahr

Entertainment Computing 43 (2022) 100500
7
their reactions are to certain scenes. Based on this assessment, the plot
moves further into a calculated direction [25]. This type of agency can
be considered involuntary, because the audience’s bodies take the upper
hand, and the audience themselves cannot deliberately manoeuvre their
way to their desired path. In this paper, this approach to agency will be
referred to as “involuntary agency.”
7. Approaches to increased agency
The purpose of this paper is to identify different approaches to
agency and assess their suitability for titles like Bandersnatch, a
nonlinear, highly restrictive, branching structure lm, in order to
improve the perceived agency. The restrictions on agency in Bander-
snatch stem from the binary nature of the choices, as well as further
restrictions that enforce a given narrative path and do not allow for
alternatives altogether. Furthermore, certain consequences of choices do
not seem to correlate with the audience’s decisions and can thus be
deemed arbitrary. Though one may refute this critique of Bandersnatch
in particular as a thematic choice used to reect the chaotic universe in
which the lm takes place, the points of critique remain valid for the
general development of a model with augmented agency for titles of that
format.
Obvious approaches to increase agency would be allowing for more
instances in which the audience can take control of the narrative and
increasing the number of controllable characters. In the interactive
drama game Heavy Rain (2010), players can make decisions on behalf of
multiple characters, although their interests are in conict as is revealed
in a plot twist in the end [26]. While Bandersnatch has a closing scene in
which a second character, namely Pearl, can be controlled, this is done
after a time jump of over 30 years, which eliminates any conicts of
interest between her and Stefan. Therefore, exploring this option in
more detail, and in particular implementing it, would signicantly in-
crease production costs and would only work well if the plotlines were
cleanly intertwined. Moreover, having control over multiple characters
can pose the threat of being too overwhelming and hence cause the
audience to lose any sense of control over the plot, which would be the
opposite of the desired effect.
Another approach which Bandersnatch may seem to have tiptoed
around – or at least reected upon – is the idea of adding quick time
events. This is seen when Stefan starts ghting his psychologist and then
his father, as the player can choose between two options that would
make him perform different attacks. Provided that the controls available
on the devices compatible with the lm allowed it, this scene could be
redesigned or even elaborated by allowing the more impactful attack
only if the player acted faster or succeeded at inputting a specic key-
combination shown on screen. In Bandersnatch, both attacks trigger
the same follow-up scene. Initially, as Bandersnatch creator Charlie
Brooke said in a podcast interview with RHLSTP, the creative team
wanted to add puzzles that audiences would have to solve to keep the
story moving forward – comparable to escape the room games. How-
ever, after conducting the player testing this idea was discarded as au-
diences did not comprehend what they were expected to do [27] –
arguably because this increased form of interaction was unfamiliar to
the target audience of Bandersnatch, which includes people who do not
engage with digital games and puzzles. Nevertheless, implementing
quick time events – even if they would have to be highly restrictive for
this format – can be an interesting feature in interactive lm, especially
if they lead to different succeeding scenarios. This can be observed in
Heavy Rain, where the player is required to navigate through the
chapters with one of the playable characters by nding clues, solving
riddles and beating quick time events. During these playable scenes, it is
up to the player’s prociency how much time is spent on them, though
hints are offered if the player has spent too much time on a challenge.
Therefore, the narrative momentum can be negatively impacted in
Heavy Rain, if a player gets stuck during a playable scene. A workaround
to diminish the decrease of narrative momentum would be to add a
timer to regulate how much time the player has to solve the task suc-
cessfully. Implementing quick time events in Bandersnatch would in-
crease interactivity in the title, however as established in section 2.2
increased (inter)activity does not correlate with increased agency.
Therefore, while this idea could be a fun addition to an interactive lm,
it would not increase agency. Although unfavourable outcomes would
be associated with failure to complete a quick time event successfully, it
is usually not the player’s choice, how well they perform at these tasks.
A more effective way to increase agency can arguably be reached by
employing one of the approaches to agency discussed in section 6.
Invisible agency can be implemented in Bandersnatch by tracking the
intentions of the audience behind all choices they make and maintaining
state of this evaluation in order to unlock suitable plotlines and endings.
In this fashion, the audience’s input would not only prompt an imme-
diate response by the controllable character, but would also determine
the later narrative path in accordance with the assessed behavioural
pattern. Looking back at Murray’s denition of agency being “the
satisfying power to take meaningful action and see the results of our
decisions and choices” [9,p. 159], we can conclude that every narrative
choice is made “consciously and visibly and the outcome is instantly
associated with it” [21,p. 180] as S
¸engün states. With an implementa-
tion of invisible agency, users would associate outcomes with their in-
puts in two separate dimensions: Primarily, the interface gives away the
choices that can be made, and their direct consequences are screened
immediately after, establishing a clear connection between input and
output. Additionally, plotlines which are unlocked further down the line
could be attributed to a cumulation of previous inputs.
The approach to agency used in Switching allows for a choice between
passive consumption and active intervention. In Bandersnatch, a similar
function is offered, as the audience can decide against making a choice
and letting the system make a default choice for them. However, due to
the interface in Bandersnatch, the choice points as well as the two options
are obvious to the audience unlike in Switching. Furthermore, the default
option is one of the two presented, which means that the logic of passive
consumption versus active intervention as seen in Switching does not
apply in Bandersnatch: Restraining oneself from making a choice (i.e.,
passive consumption) through refusal of active intervention, still results
in one of the options being played out exactly as if the audience had
actively intervened and chosen it. Therefore, in order to duplicate the
form of agency used in Switching, it would require adding a function in
Bandersnatch that lets the audience intervene or watch the scenario play
out without knowing what the consequences of either of the options are.
Switching’s solution of using visual clues performed by the actors during
which the audience can press a button to intervene could also be
implemented in Bandersnatch. Alternatively, the interface in Bander-
snatch could be used by displaying an ambiguous option such as
“Intervene” that could either be chosen or not. Although this design idea
might be interesting for titles like Bandersnatch and could be very suit-
able if it t well with the theme as it did in Switching, it would not
inherently increase agency. The vague options could trigger unforeseen
consequences, and if executed well, the audience might believe that they
are actually responsible for them, but it seems more likely that the
audience would not be able to identify with the consequences they had
brought on simply by choosing to interfere or sit back. It can be argued
that this approach to agency does not have much thematic exibility, as
it is best employed for titles, where not interfering would mean
remaining stuck in a loop – or in an unhappy relationship, like in
Switching.
Involuntary agency, like that pioneered in Obsession, is rather dif-
cult to achieve. While biosensors have become relatively common (e.g.,
Fitbits, smartwatches), their narrative efcacy requires a specic setting
that is free from external inuences that might hinder full focus and
thereby pollute the bio data. For this reason, this approach to agency
does not seem suitable for interactive lms that are designed to be
streamed and viewed in any desired setting and location.
Based on this discussion, we can classify the different approaches to
A.M. Rezk and M. Haahr

Entertainment Computing 43 (2022) 100500
8
agency regarding their exibility, starting with how compatible they are
with different devices, bearing in mind that titles that are to be released
on streaming platforms, as Bandersnatch was on Netix, have to be
compatible with a variety of different devices. We present this summary
in the form of Table 2 in which the device compatibility is broken down
into “UI Requirements,” where the visual interface as well as the
required controllers are assessed, as well as “State and Processing Re-
quirements,” to analyse which approaches to agency would require
state-tracking and background processing. These two columns help
narrow down the suitable platforms, as for instance gaming consoles
have a xed controller but congurable UI and comprehensive state and
processing capabilities, whereas DVD players, as used for Switching, have
a xed controller (remote control) and limited UI and completely lack
state and processing capabilities. Furthermore, the “Thematic Speci-
city” and the question of whether agency can be increased with these
approaches (“Increased Agency”) are further elaborated in their
Table 2
Analytic assessment of exibility of different approaches to agency.
UI Requirements
State and
Processing
Requirements
Thematic Specificity
Increased Agency
Multiple
Controllable
Characters
High flexibility
regarding the UI that
helps users
understand that they
can control multiple
characters. Devices
must be compatible
with any form of
desired interaction.
High
flexibility, as
no state-
tracking or
processing is
necessarily
needed.
Thematic specificity only
exists in the context that
users might be faced
with conflicts of interest
that they have to resolve
themselves as they are
controlling multiple
characters.
Could increase overall
agency over the
universe rather than
over one individual.
Threat of being too
overwhelming to the
audience if not
executed well.
Quick Time
Events
Required key input or
motion must be
displayed in the UI.
Devices must allow
for input of different
keys or have sensors
to track motions to
assess motor skills.
System must
be able to
process the
key inputs or
detected
motions.
Medium thematic
specificity is observed
here as this feature can
mainly be added in
action, fighting or
competition scenes.
Increased
(inter)activity does not
correlate with
increased agency.
Invisible
Agency
High flexibility: can
be implemented
based on how exactly
the system will
collect data to create
player models. Can
be done with and
without visible UI.
Devices must be
compatible with any
desired form of
interaction.
System must
be able to
track state and
process
gathered data
immediately
and as
narrative
progresses to
output
calculated
plotlines and
endings.
Player modelling can be
based on any given
criterion, therefore there
is no thematic
specificity. For example,
in Silent Hill 2, the
criterion is the
psychology of the player
because it is a
psychological horror
game.
Due to state-tracking
the user’s input
prompts immediate
responses as well as
calculated plotlines
later on, giving the
user another layer of
(invisible) agency.
Obscured
Agency
High flexibility, as
choice points can be
made visible with UI
or be implied with
narrative clues.
Devices must be
compatible with any
desired form of
interaction.
High
flexibility, as
no state-
tracking or
heavy
processing is
necessarily
needed.
If choice points are
hidden (no UI), there
would be no feedback
when interaction has
taken place. Users decide
if they intervene without
knowledge of the
consequences. To make
sense of their agency, it
is best used in looping
IDN structures, rendering
the thematic specificity
high.
Users are unaware of
the consequences of
their intervention or
refusal to intervene.
The choice points
might also be
obscured, depending
on whether or not an
interface is added.
Perceived agency is
thus not increased
unless the title
includes loops or is
replayed.
Involuntary
Agency
Low flexibility, while
no digital interface is
required, biosensors
are used as
controllers. For
accurate results, a
very specific setting
is needed.
Gathered data
from
biosensors
must be
processed.
Depending on which
physical responses are to
be measured, the only
restriction here is that the
title has to be likely to
trigger responses.
The audience is aware
that their physical
responses influence
the narrative, but they
cannot fully control
them, thus this form of
agency is
involuntary. Users are
aware of their
surrender of control.
A.M. Rezk and M. Haahr

Entertainment Computing 43 (2022) 100500
9
respective columns. This table visualises the degree of exibility of the
introduced approaches to agency regarding the relevant criteria. The
colours help distinguish between high (green), medium (yellow), and
low (red) exibility, for the rst three criteria. For the criterion
“Increased Agency,” these colours are used to visualize high, medium, or
low increase of agency.
We can conclude from the discussion and the assessment in the table
that the most promising approach to increased agency, with the least
restrictions regarding its exibility, is invisible agency. Moving forward
in this paper, invisible agency will be explored more thoroughly
regarding its implementation in Bandersnatch as well as titles with the
same format.
7.1. Benets of invisible agency in Bandersnatch-like titles
When implementing invisible agency, the player’s behavioural ten-
dencies are assessed in order to unlock suitable storylines [21]. There-
fore, it is vital to identify a criterion that is assessed whenever the
audience makes a choice. According to S¸engün, it is not advisable to base
the assessment on criteria such as ethics and morality, as options offered
might either be too obviously polarized and therefore not subtle enough,
or they may be too similar and result in a moral dilemma that in turn
creates a challenge in assessing the player’s intentions.
In Bandersnatch, a promising criterion for evaluation would be the
willingness to take risks or the propensity towards self-destructive
behaviour. These tendencies can be identied at multiple choice
points, such as when Stefan is so frustrated that he would either (a)
destroy his computer and with that all of his programming progress and
jeopardize his imminent career plans; or (b) deal his frustration by only
hitting the desk. Given the criterion, the next step is to come up with a
measuring unit for the behavioural pattern that is to be assessed,
calculating it and maintaining state throughout the title and then ulti-
mately unlocking plotlines and endings that constitute meaningful
consequences to the assessed intentions of the audience. By having later
scenes, as well as the ending, tied to the accumulated results of previous
inputs, Bandersnatch would no longer seem arbitrary in the causality of
its events. Instead, the limited agency of the interactive lm format
would be increased, as meaningful results for the audience’s input
would be observed.
This design strategy would require an overall restructuring of the
lm’s scenes in order to respond well to the user’s input. It can be
entirely up to the designers at which point the evaluated intentions and
risk-taking behaviour of the audience would bear consequences, but the
least invasive alteration would probably occur if these consequences are
displayed shortly before a potential ending is reached. The consequences
could be shown in the actions that occur in the unlocked scenes, but they
could also inuence the presented options at decision points in said
scenes.
By using invisible agency to create a more meaningful chain of
causality, the sense of agency can ultimately be enhanced, however the
oblivious audience would not notice it. Being oblivious is initially
necessary, as players who are aware that the title has invisible agency
are likely to try to manipulate the lm in a certain direction once they
know that their behaviour is being evaluated. In doing so, the unlocked
ending would no longer reect a player model of them. In this fashion,
we would argue that it adds to the level of enjoyment not to be aware of
the invisible agency when playing through the interactive lm for the
rst time. Once the players are aware of this additional layer of agency,
they might – motivated by the challenge – become more likely to replay
the lm in order to manipulate it, and this may in turn improve the per-
title replayability of Bandersnatch.
As established, Bandersnatch has a kaleidoscopic form, meaning that
audiences are encouraged to explore alternative paths upon reaching an
ofcial ending, which increases the per-scene replayability. Moving
beyond this, it is also important to note that Bandersnatch relies on
dramatic compression to accelerate the arrival at a decisive choice point,
after “respawning,” i.e., exploring a variation of the plot. This means
that the audience does not have to sit through a repetition of previous
scenes in order to make a change within the same playthrough. While
this feature might increase the attractiveness of exploring multiple paths
in one sitting, it is a double-edged sword. The constant state-tracking
required for invisible agency could constrain the audience from
unlocking all narrative paths in one singular playthrough, since the
paths are required to match the created player models.
It is therefore worth exploring how the kaleidoscopic design can
coexist with invisible agency and the generated player model. Broadly
speaking, there are two options to facilitate backtracking, the rst being
to reset the player model either entirely, or at least discarding all in-
formation that was gathered after the choice point that the player is
respawned into. The benet of this option is that it would not limit the
potential storylines that can be unlocked as severely (or even at all),
because some (if not all) data used for the player model will be over-
written. The drawback, however, is that the player model could become
much less accurate if gathered information is being replaced so easily.
The second option is to accumulate all data about the player during one
playthrough even when backtracking is involved, meaning no data is
lost. Doing so also allows the system to assess increased instances of user
behaviour, such as the audience’s choice to engage with the title longer,
perhaps out of curiosity, lack of perceived closure or satisfaction with
the reached ending. This information can then be used to further adjust
the alternatively unlocked plotlines and endings. The main drawback of
this approach is that by retaining all gathered data rather than over-
writing some of it, the audience can be predestined to a limited number
of alternative storylines early on, as only those would match their
overall player models. A potential solution for this design limitation
would be screening the most suitable plotlines rst and gradually going
over to the less suitable ones.
In order to facilitate a well-functioning form of invisible agency, the
juxtaposed options at each decision point should not be obvious choices
on opposite sides of a spectrum, as this would take away any possible
challenge for the aware audience. By evaluating the risk-taking behav-
iour as a calculation of percentages rather than as a simple binary
branching at each decision point, the assessment can be conducted in a
more sophisticated manner. These percentages can then be accumulated
to calculate the right outcome just before it is to be screened.
It is also worth exploring the differences in format, duration and
pacing. Silent Hill 2 is an over eight-hour long gaming experience in
which the user’s behaviour determines the possible endings, while
Bandersnatch is a signicantly shorter interactive lm of about one and a
half to two and a half hours in length. However, Bandersnatch asks its
audience to make concrete, binary choices, rather than gathering data
from the way the user chooses to interact with the story world as done in
Silent Hill 2. Therefore, though being a fast-paced title in comparison,
Bandersnatch has the tools to assess certain characteristics in the audi-
ence more efciently without its overall duration becoming a threat to
the accuracy of the assessment.
Ultimately, the narrative momentum as previously discussed cannot
be inuenced in Bandersnatch’s format as the audience has to make a
decision within a dened amount of time even if the features of invisible
agency were to be added.
7.2. Discussion of authoring tools
Interactive lms can be authored using a variety of tools, many of
which are open source. Commonly used software and languages are
Twine, Ink, Inform, and StoryPlaces, which can be used for different
results and functionalities. For the prototyping process of Bandersnatch,
the authoring tool Twine was used. Twine is a visual hypertext devel-
opment tool which does not require programming knowledge for its
basic functionalities. To add more complexity to interactive experiences
creators can use one of the supported story formats (similar to scripting
languages), such as SugarCube, Harlowe, and Snowman.
A.M. Rezk and M. Haahr

Entertainment Computing 43 (2022) 100500
10
Referring to the implementation of invisible agency, the use of Twine
as a prototyping software would still be possible, by for instance using
variables to maintain state and adding conditional statements to deter-
mine which scenes are to be unlocked under which circumstances.
For Bandersnatch in particular, Netix programmed its own software
that according to show creator Charlie Brooker made it possible to
import the code alongside the respective scenes, allowing for a smooth
transition between the prototyping/authoring and production stage
[27].
7.3. Importance of the awareness of agency
After establishing that invisible agency expects rst-time users to be
unaware of the player modelling feature, a question arises about the
importance of the awareness of agency. Can the audience feel in control
if they are oblivious of their inuence? This is where a line has to be
drawn between the two terms agency and control.
Having agency does not mean being in control; it merely means that
“meaningful” outputs will be achieved in a dynamic system. The word
“meaningful” is subject to the logic of said system, i.e., the logic of the
world in which the story takes place. Therefore, the denition of the
word “meaningful” must bend to the logic of the relevant story world
and for a satisfying experience also make thematic sense. While many
narrative outputs will be hardly predictable, as would be expected from
any form of narrative, this does not diminish the level of agency the
users have, even if they might feel like they were not in control. The
pleasure of agency, as Murray stated, comes from “making something
happen in a dynamically responsive world” [12]. It can hence be dened
as the synergy of affecting the plot and in turn being affected by the
overall experience [28]. This quality of affecting is not the same as being
in control. The audience might be holding the strings of the controllable
character(s), but this power does not translate to external forces, thereby
allowing for unpredictability in the narrative, which goes against the
notion of being in control.
With invisible agency, plotlines that match a certain player model of
the user will be unlocked as the plot progresses. This creates a second
layer through which the user can affect the plot, as their input no longer
prompts only immediate responses. First-time users will likely be ex-
pected to only be aware of their power to manoeuvre the immediate
responses whilst remaining unaware of their inuence on the plot on a
larger scale.
As agency cannot be equated with being in control, it is worthwhile
to establish whether there even is a necessity for the scale of agency to be
recognizable up front. Users who nd out that the narrative experience
they have achieved was adjusted through a modelling of them will un-
derstand that their choices affected the plot more than they initially
thought. This in turn will make them understand that they have more
agency than expected and as elaborated in section 7.1, might prompt
them to replay the entire title, this time knowing that their choices have
a more signicant narrative effect.
This leads to the conclusion that invisible agency while not being
perceivable for rst-time users does not risk rendering the concept
ineffective, as long as the system reveals this additional feature at the
end of the experience, so users can become aware of their multi-faceted
level of agency.
8. Conclusion
This paper has identied several different approaches to agency,
which, if implemented for a suitable interactive title, can increase the
user’s perceived agency. For highly restrictive, branching structure
lms, like Bandersnatch, we found that by adding features of invisible
agency, the overall perceived agency among the users can be amplied
without requiring highly invasive modications to the title. The narra-
tive momentum would not be jeopardized, as the UI would remain un-
changed, meaning choices would still be made within the allotted time.
Furthermore, the additional layer of invisible agency could potentially
lead to audiences replaying the title, as they are aiming for alternative
plotlines: Provided that the audiences of interactive lms can be
assumed generally to be aiming for a certain narrative outcome, adding
the layer of invisible agency would offer an additional challenge for the
audience – especially if they are replaying the entire lm or are aware of
the invisible agency feature – as they would try to manoeuvre the
choices carefully to get their desired result. In combination with this
challenge, this solution could prove itself promising in navigating the
ne line between narrative and game successfully and invisibly. Finally,
implementing invisible agency in interactive lms like Bandersnatch can
contribute to the creation of more meaningful sequences of events in the
instantiated narratives in accordance with the audience’s input, and
thereby the overall felt agency or “free will” would be increased.
To conclude, while scarcely explored in the past, invisible agency
could have the potential to elevate IDNs to more captivating and
engaging titles with a higher degree of perceived agency. However, this
design suggestion inevitably raises the question of what will be done
with the data gathered from the assessment of the audience’s behaviour.
Streaming platforms such as Netix and Prime Video track a consider-
able range of user behaviours, for instance for their personalised
recommendation systems [29,30]. For IDNs, especially those that rely
heavily on state-tracking and processing, as would be the case if invisible
agency was to be added, the in-title-behaviour of audiences would likely
give the algorithm even more data about the users. Technology policy
researcher Michael Veale used GDPR to formally ask Netix to share the
saved data regarding the choices he made in Bandersnatch and found
that the streaming platform stored every single user input without
indicating for how long this data would be stored [31]. Taking a step
back, it is worth questioning if this further envelopment of audiences
into the datacation of interactive entertainment is a desirable devel-
opment, an inevitable side-effect, or something that should rather be
prevented for the sake of the media format. This question could open the
door to a more elaborate discussion regarding the ethical and legal
grounds of data collection of IDNs through behaviour-assessments for
monetisation purposes. Another important question to consider here, is
if users who are aware of the invisible agency component and know that
they are being proled might be deterred from engaging with such IDNs.
Moreover, to further develop this suggested approach to increased
agency, it would also be imperative to have actual user studies on
invisible agency in interactive lms, rather than theoretical assump-
tions, to be able to steer this promising design idea into the right di-
rection in the long run.
Based on the analytical assessment of invisible agency, we can
conclude that this concept is profoundly adaptable and can therefore be
applicable to interactive stories with different themes. Invisible agency
also offers a high degree of design freedom in terms of what the player
model should be based on. For example, in the hypothetical example of
adding invisible agency to Bandersnatch, we identied the potential to
model players based on their risk-taking behaviour and tendency to be
self-destructive. Moving beyond this, we recognize a potential challenge
to balance the coexistence between the kaleidoscopic design and the
generated player model through invisible agency. However, there are
different design suggestions to increase replayability through back-
tracking whilst maintaining the player model, which could be elabo-
rated, tried, and tested in the future. Nevertheless, it remains clear that
due to the duality of its effect, invisible agency can heighten the feeling
of “making something happen in a dynamically responsive world” [12]
and therefore be a powerful tool in interactive storytelling.
Declaration of Competing Interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
A.M. Rezk and M. Haahr

Entertainment Computing 43 (2022) 100500
11
References
[1] R. Donaghey, Netix Is Going to Make a Lot More Weird, Interactive Movies Likes
‘Bandersnatch’. Vice Magazine [Online] 2019. https://www.vice.com/en/article
/7xnq3a/netix-more-interactive-movies-like-bandersnatch-rom-coms-todd-yellin-
interview-vgtrn (accessed Dec 15, 2020).
[2] C. Watson, Kimmy vs. the reverend, this is the new interactive movie from Netix.
Freegame Tips [Online] 2020. https://freegametips.com/kimmy-vs-the-reverend-
this-is-the-new-interactive-movie-from-netix/ (accessed Jan 12, 2021).
[3] CrtlMovie. https://www.ctrlmovie.com (accessed Dec 15, 2020).
[4] A.M. Rezk, M. Haahr, The case for invisibility: understanding and improving
agency in Black Mirror’s Bandersnatch and other interactive digital narrative
works. In: Interactive Storytelling, Proceedings of the 13th International
Conference on Interactive Digital Storytelling, ICIDS 2020, Bournemouth (UK),
November 3-6, 2020; A. G. Bosser, D. E. Millard, C. Hargood, Eds.; Springer: Cham,
2020; pp 178-189. https://doi.org/10.1007/978-3-030-62516-0.
[5] H. Koenitz, Towards a Specic Theory of Interactive Digital Narrative. In Interactive
Digital Narrative, 1st ed.; H. Koenitz, G. Ferri, M. Haahr, D. Sezen, T. I˙. Sezen, Eds.;
Routledge, 2015; pp 96-101. DOI: 10.4324/9781315769189.
[6] C. Roth, T. van Nuenen, H. Koenitz, Ludonarrative Hermeneutics: A Way out and
the Narrative Paradox. In Interactive Storytelling, Proceedings of the 11th
International Conference on Interactive Digital Storytelling, ICIDS 2018, Dublin,
December 5–8, 2018; R. Rouse, H. Koenitz, M. Haahr, Eds.; Springer: Cham, 2018;
pp 93-106. DOI: 10.1007/978-3-030-04028-4.
[7] D. Fox Harrell, J. Zhu, Agency Play: Dimensions of Agency for Interactive Narrative
Design, in: Intelligent Narrative Technologies II, The AAAI Press, 2009, pp. 44–52.
[8] M. L. Ryan, From Narrative Games to Playable Stories: Toward a Poetics of
Interactive Narrative. In: Storyworlds: A Journal of Narrative Studies, 2009, Vol. 1,
pp 43-57. https://www.jstor.org/stable/25663007 (accessed Dec 15, 2020).
[9] J.H. Murray, Hamlet on Holodeck: The Future of Narrative in Cyberspace, 2nd ed.,
MIT Press, Cambridge and London, 2017.
[10] N. Knoller, U. Ben-Arie, The Holodeck is all Around Us–Interface Dispositifs in
Interactive Digital Storytelling. In Interactive Digital Narrative, 1st ed.; H. Koenitz,
G. Ferri, M. Haahr, D. Sezen, T. I˙. Sezen, Eds.; Routledge, 2015; pp 51-66. DOI:
10.4324/9781315769189.
[11] G. N. Yannakakis, P. Spronck, D. Loiacono, E. Andr´
e, Player Modeling. In Articial
and Computational Intelligence in Games; S. M. Lucas, M. Mateas, M. Preuss, P.
Spronck, J. Togelius, Eds.; Dagstuhl Publishing, 2013; pp 45-60. DOI: 10.4230/
DFU.Vol6.12191.45.
[12] J. H. Murray, Janet H. Murray Humanistic Design for Emerging Medium. Glossary.
https://inventingthemedium.com/glossary (accessed Dec 28, 2020).
[13] S. K. Ashwell, Standard Patterns in Choice-Based Games. These Heterogenous Tasks
[Online] 2015. https://heterogenoustasks.wordpress.com/2015/01/26/standard-
patterns-in-choice-based-games/ (accessed Dec 18, 2020).
[14] C. Roth, H. Koenitz, Bandersnatch, Yea or Nay? Reception and User Experience of
an Interactive Digital Narrative Video. In TCX’19, Proceedings of the 2019 ACM
International Conference on Interactive Experiences for TV and Online Video,
Salford (UK), June 5-7, 2019, Association for Computing Machinery, 2019; pp.
247-254.
[15] J. H. Murray, Research into Interactive Digital Narrative: A Kaleidoscopic View. In
Interactive Storytelling, Proceedings of the 11th International Conference on
Interactive Digital Storytelling, ICIDS 2018, Dublin, December 5–8, 2018; R.
Rouse, H. Koenitz, M. Haahr, Eds.; Springer: Cham, 2018; pp 3-17. DOI: 10.1007/
978-3-030-04028-4.
[16] A. Mitchell, L. Kway, “How do I Restart This Thing?” Repeat Experience and
Resistance to Closure in Rewind Storygames. In Interactive Storytelling, Proceedings
of the 13th International Conference on Interactive Digital Storytelling, ICIDS
2020, Bournemouth (UK), November 3-6, 2020; A. G. Bosser, D. E. Millard, C.
Hargood, Eds.; Springer: Cham, 2020; pp 164-177. DOI: 10.1007/978-3-030-
62516-0.
[17] J. H. Murray, Janet Murray responds in turn, Electronic Book Review [Online] 2004.
http://electronicbookreview.com/essay/janet-murray-responds-in-turn/ (accessed
Feb 2, 2021).
[18] J. H. Murray, Why Paris Needs Hector and Lancelot Needs Mordred: Using
Traditional Narrative Roles and Functions for Dramatic Compression in Interactive
Narrative. In Interactive Storytelling, Proceedings of the 4th International
Conference on Interactive Digital Storytelling, ICIDS 2011, Vancouver, November
28 – December 1, 2011; M. Si, D. Thue, E. Andr´
e, J. C. Lester, J. Tanenbaum, V.
Zammitto, Eds.; Springer: Cham, 2011; pp 13-24. DOI: 10.1007/978-3-642-25289-
1.
[19] E. Short, Bandersnatch (Netix). Emily Short’s Interactive Storytelling [Online] 2019.
https://emshort.blog/2019/02/19/39189/ (accessed Dec 15, 2020).
[20] L. Kohlhoff, F. Nack, How Relevant Is Your Choice?: User Engagement and
Perceived Agency in Interactive Digital Narratives on Video Streaming Platforms.
In Interactive Storytelling, Proceedings of the 12th International Conference on
Interactive Digital Storytelling, ICIDS 2019, Little Cottonwood Canyon (USA),
November 19–22, 2019; R. E., Cardona-Rivera, A. Sullivan, R. M. Young, Eds.;
Springer: Cham, 2019; pp 73-85. DOI: 10.1007/978-3-030-33894-7.
[21] S. S
¸engün, Silent Hill and Curious Case of Invisible Agency. In Interactive
Storytelling, Proceedings of the 6th International Conference, ICIDS 2013, Istanbul,
November 6–9, 2013; H. Koenitz, T. I˙. Sezen, G. Ferri, M. Haahr, D. Sezen, G.
Çatak, Eds.; Springer: Cham, 2013; pp 180-185. DOI: 10.1007/978-3-319-02756-2.
[22] A. Romero, WTF is the Pax Demon in Black Mirror Bandersnatch? Renery29
[Online] 2018. https://www.renery29.com/en-us/2018/12/220437/pax-demon-
black-mirror-bandersnatch-netix-real-pac (accessed Jan 29, 2021).
[23] B. Makedonski, 9 out of 10 people played Mass Effect as Paragon. Destructoid
[Online] 2020. https://www.destructoid.com/9-out-of-10-people-played-mass-
effect-as-paragon-581278.phtml (accessed Dec 27, 2020).
[24] Switching. https://web.archive.org/web/20201127234605/http://www.
switching.dk/en/ (accessed Jan 2, 2021 through Wayback Machine’s calendar
entry of Nov 27, 2020).
[25] Enactive Cinema. http://www.enactivecinema.net (accessed Dec 29, 2020).
[26] L. Kietzmann, Interview: Spoiling Heavy Rain with David Cage. Engadget [Online]
2010. https://www.engadget.com/2010/03/19/interview-spoiling-heavy-rain-
with-david-cage/ (acccessed Dec 18, 2021).
[27] R. Herring, RHLSTP 268 - Charlie Brooker. RHLSTP [Podcast] 2020. https://play.
acast.com/s/rhlstp/rhlstp268-charliebrooker (accessed Nov 1, 2021).
[28] H. Koenitz, What Game Narrative Are We Talking About? An Ontological Mapping
of the Foundational Canon of Interactive Narrative Forms. In: Gaming and the Arts
of Storytelling; Special Issue, D. Jayemanne, Ed.; Arts, 2019; pp. 4-14. DOI:
10.3390/arts7040051.
[29] Netix Help Center. https://help.netix.com/en/node/100639
(accessed Jan 28, 2020).
[30] D. Chong, Deep Dive into Netix’s Recommender System. Towards Data Science
[Online] 2020. https://towardsdatascience.com/deep-dive-into-netixs-
recommender-system-341806ae3b48 (accessed Feb 5, 2021).
[31] M. Gault, Netix Has Saved Every Choice You’ve Ever Made in ‘Black Mirror:
Bandersnatch’, Vice Magazine [Online] 2019. https://www.vice.com/en/article/
j57gkk/netix-has-saved-every-choice-youve-ever-made-in-black-mirror-
bandersnatch (accessed Feb 3, 2021).
A.M. Rezk and M. Haahr