Source: docs/agent-briefs.md
Agent Briefs
An agent brief turns a selected solution into bounded execution.
It carries aim, mechanism, feedback, guardrails, scope, checks, and stop conditions in one artifact.
Learn
A useful brief contains:
- aim;
- selected problem statement;
- selected solution level;
- rejected solution levels;
- mechanism;
- feedback signal;
- guardrails;
- context to inspect first;
- behavior contract;
- commands and checks;
- explicit non-goals;
- stop conditions;
- review criteria.
The brief should make it possible to reject a plausible patch.
Practice
Use templates/agent-brief.md after /solution-space and the evidence checklist.
Before giving it to /execute, ask:
- Does every tactic connect to the aim?
- Is the causal mechanism stated plainly?
- Are rejected solution levels named?
- Are checks tied to behavior, not implementation details?
- Does the agent know where to stop instead of guessing?
- Could a reviewer use this brief without reading the whole conversation?
Artifact
Use templates/agent-brief.md. The brief is an execution contract, not an implementation wish list. It must preserve this shape:
| Section | Purpose |
|---|---|
| Purpose and aim | why this slice matters |
| Problem statement | the selected framing |
| Selected and rejected solution levels | what level to build and what not to build |
| Mechanism, feedback, guardrails | why it should work, how to check it, what must not break |
| Context to inspect first | where the agent should start reading |
| Behavior contract | required behavior and non-goals |
| Commands and acceptance checks | external verification |
| Old behavior the checks should catch | the regression that prevents shallow success |
| Stop conditions | when to stop instead of guessing |
| Review checklist | how a reviewer can reject the work |
The brief is consumed by /execute, /review, /dissent, and knowledge extraction. A reviewer should be able to reject the work using only this artifact plus the diff; otherwise, the brief lacks enough specificity.
Review check
Reject a brief if:
- it only says what to edit;
- it omits mechanism;
- it hides why other solution levels were rejected;
- it lacks commands or acceptance checks;
- it does not name non-goals;
- it gives the agent permission to expand scope silently.
Go deeper
templates/agent-brief.md— handoff template.docs/strategy-clarity.md— aim, mechanism, feedback, guardrail, and solution level.- Alignment Is the Constraint — why speed without alignment creates more wrong work faster.
- Documenting Strategy — why artifacts should connect vision, needs, strategy, tactics, and evidence.
Navigation
- Previous: Evidence and Evals
- Up: Docs Home / Curriculum
- Next: Strategy Clarity