This report describes research conducted during 1989 and 1990 on the cognitive characteristics of a corpus of anesthesia critical incidents. The incidents were collected by monitoring and transcribing the regular quality assurance conferences in a large, university anesthesiology department. The 57 reports of incidents were analyzed by constructing protocols which traced the flow of attention and the knowledge activation sequence of the participants. Characteristics of the resulting protocols were used to divide the collection into five categories: acute incidents, going sour incidents, inevitable outcome incidents, airway incidents, and non-incident incidents. Of these, the acute and going sour categories represent distinct forms of incident evolution. The implications of this distinction are discussed in the report. Nearly all of the incidents involve human cognitive performance features. Cognition clearly plays a role in avoiding incidents but also in aborting and recovering from incidents in progress. Moreover, it is clear that subtle variations in cognitive function may playa crucial role in anesthetic disasters, of which incidents are taken to be prototypes. Review of the corpus reveals the different cognitive functions involved in anesthesia and anesthesia incidents. These cover a wide range including classic aspects of cognition, for example the direction of attention, and complex and poorly understood aspects such as situation awareness. The cognitive features include dealing with competing goals, dealing with competing indicators, the limitations of imperfect models, knowledge activation failures, the role of learned procedures and assumptions in reducing cognitive workload, failure to integrate multiple themes, organizational factors, and planning. These presence of these different cognitive features and cognitive failures in a single discipline is significant because it enhances and supports separate findings from other domains (e.g. nuclear power plant operation, commercial aviation) and also because it provides strong support for the contention that operators acting in these semantically complex, time pressured, high consequence domains face common problems and adopt similar strategies for dealing with them. The report demonstrates the way in which cognitive analysis of incidents can be accomplished in anesthesia and in other domains and suggests a system for categorizing the results obtained. It also raises questions about the adequacy of evaluations of risk and safety that do not explicitly account for the cognitive aspects of incidents and their evolution. In order to make real progress on safety in domains that depend critically on human operators it is necessary to examine and assess human cognitive performance, a process which requires large amounts of data and careful reconstruction. Such cognitive analysis is difficult. It requires substantial experience, skill, and effort and depends on acquiring and sifting through large quantities of data. This should not be suprising, since the domain itself is one characterized by experience, skill, effort, and large quantities of data. The challenge for us and for other researchers is to perform more such analyses and extend and refine the techniques described here and to link the analyses to those from other domains.