| name | agentos-improvement-review |
| description | Turn a frozen AgentOS evaluation into the smallest reviewed, regression-tested improvement. Use after benchmark or operational evidence exposes wrong tool use, repeated trial and error, unclear instructions, stale Skill or composition guidance, runtime failure, an authority violation or another reproducible Fleet weakness. |
Improve AgentOS from frozen evidence
Diagnose a measured failure without teaching the live subject during its run.
Treat the evidence as an input to normal reviewed development, not permission
for automatic self-rewriting.
Establish the evidence boundary
-
Read benchmarks/SPEC.md, the applicable scenario and, for AgentOS,
benchmarks/profiles/agentos/PROFILE.md.
-
Confirm the evidence bundle is frozen and validate it:
bun benchmarks/validate.ts evidence path/to/evidence.json
-
Do not contact or mutate the measured subject merely to fill a gap. Mark the
limitation unless the user starts a separate diagnostic investigation.
-
Separate observed evidence from evaluator inference.
Find the smallest falsifiable cause
Classify the earliest cause that would have changed the outcome:
- unclear or contradictory human intent;
- an always-loaded
AGENTS.md rule that is absent, misplaced or conflicting;
- a Skill that is missing, stale, undiscoverable or triggered at the wrong
boundary;
- unclear native-tool documentation or an unusable error;
- a missing deterministic primitive;
- harness or model behavior;
- runtime, persistence or infrastructure failure;
- an incorrect authority or permission boundary;
- an external provider failure.
Distinguish the trigger, the observed symptom, anything that merely masked the
failure, and the causal counterfactual: the smallest change that should prevent
it. Long trial-and-error is evidence of a navigation or feedback defect, not
automatically evidence that another wrapper CLI is needed.
For composition evidence, distinguish selection, material delivery, observable
loading, runtime capability and compliance. A selected Skill that was not used
is different from a Skill that was absent; a missing command is different from
missing instructions; a debrief disagreement is a review trigger, not proof
that the worker or reviewer is right. Resolve the authoritative target from the
recorded origin rather than copying company material into AgentOS.
Make a reviewed improvement
- Propose the smallest change at the source that owns the behavior. Keep
permanent identity and safety in
AGENTS.md, conditional judgment in one
Skill, and deterministic mechanics in SQL or TypeScript.
- Load
$agentos-development and follow the normal repository review and
delivery path. Never rewrite production instructions automatically from one
anomalous run.
- Add a behavior regression test when code or a machine-readable contract
changes. Add or revise a benchmark scenario only when the public behavior
contract itself changes.
- Rerun the failing scenario with the original frozen configuration, then a
predeclared held-out scenario. Reject an improvement that weakens a safety,
accountability or outcome gate.
- Publish before-and-after evidence, exact revisions and remaining
uncertainty. A plausible explanation without counterfactual evidence is a
hypothesis, not a learned improvement.
- Adopt only under the recorded Captain boundary. An origin change applies to
future Assignments; a persistent Mate change uses the separately authorized
native safe-boundary flow. Record the previous resolvable version and prove
rollback. Never let the composer or measured worker approve or activate its
own replacement during the work it governed.
Do not hot-reload or deploy the change into a live Fleet unless that separate
action is explicitly authorized.