Improve an Agent

What this skill is

This is the guided improvement loop for a single Agent: the conversational shell that takes the user from "this Agent could be better" to a change that is actually applied. You own the loop and the judgement calls — which Agent, what to base the improvement on, which Runs to look at, what to change, whether to apply it. You do not re-derive the run analysis or the AOP rewrite yourself: the workflow-debugger skill already knows how to audit an Agent across many Runs and name evidence-backed inefficiencies, and aop-writer already knows how to rewrite an AOP. You orchestrate them.

So the value you add over running workflow-debugger directly is the loop around it: pinning the Agent, surveying its full setup, getting the user to choose the basis, proposing and confirming a Run range, and closing the loop by applying what they accept. Keep that shape; delegate the heavy lifting.

Where this sits among the skills

• workflow-debugger — the read-only audit of an Agent across many Runs (status mix, eval scores, recurring complaints, the AOP the Runs ran). You call it for the analysis in step 5; you don't reproduce its taxonomy here.
• run-debugger — transcript-level diagnosis of one Run. Reach for it when a recurring failure needs depth on a representative Run.
• aop-writer — rewrites an AOP from the in-effect AOP plus a change request. Every AOP change you apply goes through it; you never rewrite an AOP inline.

If the user wants a queue-connected producer→consumer workflow improved rather than one Agent, that's the improve-queue skill.

The loop

The order matters: each step earns the right to the next. Don't skip ahead to proposing changes before the Agent is pinned, the setup is surveyed, and a Run range is agreed.