Skip to main content
Use native optimize when you have a declared search space, one lever to vary, and a stop rule. It is a measured loop, not permission to broaden a prompt, skill, or task corpus until a result looks better.

​
Say this

​
What happens

Your agent runs a bounded sequence: candidate, evaluate, check in the result, choose one next mutation, then repeat only while the declared rules allow it. Have it state expected time and spend before launch. Each change needs a one-lever hypothesis and a record of rejected levers. Stop when evidence is invalid or inspect-only, improvement stops, the search space is exhausted, or the next move crosses the approved spend boundary. The agent should not widen the lever, task slice, or decision claim without asking. A finalist still needs matching holdout approval before promotion or any public claim. Screening evidence is useful for iteration; it does not silently become holdout evidence.

​
What a good answer sounds like

The report should state the current candidate, the one changed lever, its hypothesis and result, the stop-rule status, and spend so far — every value from the actual Trial Result. It should say explicitly when a holdout is still required rather than implying that a screening result authorizes promotion. Ask: “Which lever changed this round?” “What is the retained denominator?” “Does the finalist still require holdout evidence?”