Agents & procedures

Operations

Input	Derived object	Activation boundary
Conversations and trajectory observations	Searchable episodes with exact source evidence	Current sources and authorization checked on read
Runtime events	Redacted normalized trajectory	Immediate record
Trace exports	OpenAI, Anthropic, or OTLP normalized steps	Unknown outcomes remain unknown
Trajectory corpus	Direct-follow graph, variants, rework, invariants	Caller reviews findings
Model-proposed intent	Evidence-bound intent candidate	Independent semantic review
Successful trajectories	Procedure candidate	Held-out gates and review
Reflections	Skillbook mutation proposal	Manual promotion
Reviewed procedure	Cached workflow match	Fresh and authorized dependencies required

Trajectory-to-procedure example

Observation	Derived result
Two successful runs share lookup_policy → issue_refund	Procedure candidate contains both stable steps
Failed run exits after issue_refund	Failure retained as negative evidence
Candidate improves score and leaks unauthorized context	Hard gate rejects promotion
Candidate passes all gates	Submitted for review. Activation requires an application commit

• Conversation knowledge
• Email conversation events
• Trajectories and agent evaluation
• Trajectory mining
• Intent mining
• Procedural learning
• Procedural knowledge