CONCEPT · DECISION
Decision
The auditable artifact produced every time an agent thinks.
What it captures
Prompt, model, latency, input tokens, output tokens, confidence score, structured output, and any human override. It is not a log line — it is a first-class record with schema.
Why it matters
Regulators, auditors, and future-you all need to answer "why did the system do that?" A decision record makes that a query, not an archaeology project.
Replay
Any decision can be re-run against a different model to compare cost, latency, and answer — the basis for the roadmap's Decision Drawer.
Shape
| Field | Type | Note |
|---|---|---|
| agentId | string | Which agent produced this decision. |
| prompt | string | Full prompt sent to the model. |
| model | string | e.g. google/gemini-3-flash-preview |
| confidence | number | 0–1 self-reported confidence. |
| override | string? | Human override, if any. |