Contribute

Improve from evidence

Turn one frozen failure into the smallest reviewed and regression-tested change.

Start an improvement review only after the evaluation bundle is frozen. Measurement and mutation are separate runs.

The review identifies one reproducible weakness: wrong tool use, repeated trial and error, stale instructions, runtime failure, authority violation or another observable Fleet defect. It finds the smallest owning source and changes that source of truth rather than appending the same rule to several docs.

Improvement loop

  1. reproduce the failure from frozen evidence;
  2. name the violated contract and smallest cause;
  3. write a regression test through the public behavior;
  4. change the owning SQL, TypeScript, Skill or instruction;
  5. review the diff and rerun the original plus held-out scenario;
  6. publish evidence including remaining limitations.

Memory can suggest a pattern, but shared behavior changes require this reviewed path. A new wrapper or daemon is not the default fix when an existing native interface or source contract owns the problem.

Last updated on

Evaluate an organization

Freeze sanitized evidence against the public AgentOS benchmark before improving it.

Release model

Build immutable artifacts from a clean exact revision and publish only after verification.

On this page

Improvement loop