Situational Awareness overview

What SA is Situational Awareness is Cerevisor's continuously-updated, three-level picture of a workflow run. The pillar sits on top of memory, the orchestrator, and persistence; it does not own state, it composes signals already produced by the rest of the harness into typed snapshots that the chat builder, the orchestrator, and the renderer can all read. The pillar is grounded in three pieces of academic work: Endsley's 3-level SA : perception (what's happening), comprehension (what it means), projection (what's likely next). Endsley 2023's human-AI SA split : when humans and AI work together, awareness divides into taskwork, agent, and teamwork dimensions. Distributed SA : awareness lives across people, tools, and artifacts, not inside any single component. In Cerevisor, this turns into: snapshots that name uncertainty explicitly, attribute every load-bearing claim to its source, and surface gaps instead of papering over them. The three slices Every situation snapshot has three slices. The chat builder and the run drawer surface them as separate panels. 1. Taskwork SA: awareness of the task What this workflow is, what's changed since you last interacted, what other workflows in the workspace you've touched. sinceLastSession : how many runs happened since you last opened the chat builder for this workflow, how many succeeded, how many didn't, when the most recent run was, and up to three newly-detected blind spots. siblingWorkflows : other workflows in the same workspace with their last-run timestamp and status. The chat-builder opener uses this slice to lead with concrete deltas instead of "what would you like to build?". 2. Agent SA: awareness of the agents What the harness believes about its own agents' recent uncertainty and the patterns it thinks hold. lastUncertaintyByRole : the most recent uncertainty snapshot for each role that has run, capped to keep the panel readable. topPatterns : patterns the harness believes hold (e.g. "researcher does better with three sub-questions than five"), each tagged with its source ( user-explicit , user-implicit , harness-inferred , harness-extrapolated ) and a confidence value. openBlindSpots : known unknowns the harness is currently tracking. This is the panel you read when something feels off. If an agent ran with verifierState: 'rejected' or no confidence marker at all, that shows up here before you have to dig through audit logs. 3. Teamwork SA: awareness of the working relationship What you and the harness have agreed on, where you disagree, and what's still an open question. openDisagreements : places where you claimed one thing and the harness's observations suggest another. Each item carries your claim, the harness's observation, and the confidence gap. conversationState : when a chat-builder session is active, the running accumulator of what you've accepted, rejected, clarified, and the questions still on the table. This is the slice that prevents the harness from suggesting the same thing again

Back to docs