Anthropic provides
Sessions hold the agent's state and memory, and the platform traces every run — tool calls, decisions, and failure modes are inspectable in the Claude Console. Active runtime is metered per session-hour.
The same concept exists locally: Claude Code saves every session as a transcript file on the machine it ran on — resumable, branchable, exportable, with hooks that fire when a session ends. Two venues, one idea: the session is the unit of conduct.
Skoor adds
Tracing is observability: the operator inspecting its own runtime. Skoor reads the same conduct from outside and checks it against policy — did the agent stay inside its scopes, was drift caught inside the committed window, is the trail complete. That is what an Agent Skoor is computed from, continuously, per agent.
Because sessions are already the metered unit of managed agents, they are the natural denominator for a behaviour score: conduct per session, accumulated into a number someone can stand behind.
On the local side the journal already exists: the transcript on your own disk is the conduct record, and a session-end hook can archive it to your audit sink before retention cleanup removes it. Governance detail worth knowing: a branched session gets its own identity — its own row — and a resumed session does not silently regain bypass permissions; the platform's own defaults lean the right way.
Stated plainly
Skoor reads the record to compute a score and holds no copy of the records the trail describes.
The assessment establishes what your agents can reach at every edge — and produces your baseline Company Skoor.