Operating
How repeated work improves
Repeated work improves when the team keeps the context, method, evidence, and correction together.
- 01A defined job creates a completion recordThe run has a stated outcome and an acceptance test.
- 02Good source context reduces guessingLess unnecessary search, fewer wrong assumptions.
- 03A Skill preserves a method that workedThe procedure survives past the person who ran it.
- 04Verification creates a reliable signalChecks tell you what actually worked rather than what looked finished.
- 05Review separates errors from preferencesAnd both of those from policy changes and one-time exceptions.
- 06A scoped update improves one layerKnowledge, the Skill, a validator, routing, or a tool rule.
- 07The next run tests whether the update helpedImprovement is measured rather than assumed.
The system does not treat every edit as a universal lesson, and experimental guidance does not change live work directly. Updates stay scoped, reviewed, versioned, and reversible.
Why this is the safety story too
A system where production agents can rewrite their own policy improves quickly and unpredictably. Keeping promotion behind human approval is what makes the learning something you can leave switched on.
Reviewed work produces candidate changes to Skills, Memory, routing, tools, and evaluators. Those candidates are scoped, tested, approved, and reversible, so a change that turns out to be wrong is a change you can undo rather than a behavior you have to diagnose.