Have it compare to the real work you've done in a sample of your previous sessions. Agents are good at simulating completing the work following newer harnesses, heuristics etc. and comparing time, tokens, quality, pitfalls.