lab: ai-factory-planner / skills / execution trace
Context: AI Factory should not be just a chat that returns text, but a working loop: accept a goal, split it into steps, choose tools and leave a readable execution trace.
Problem: an agent without a plan quickly becomes noise. It does extra work, loses context, fails to explain why a skill was chosen and may continue confidently where a human decision is required.
Approach: the planner should store state: goal, assumptions, steps, status, selected skill, input data, result and stop reason. Risky operations need explicit approval, uncertainty needs a question.
Conclusion: a good planning loop does not make AI omnipotent. It makes work observable: what has been checked, where the risk is, what can be repeated and which step produced the result.
Next step: add plan scoring: cost, risk, external-data dependency and confidence in the next action.