3. Optimization loop

The main optimization loop. The agent proposes a fast variant each round, runs it via zyme run, compares against baseline + concordance gate, and either accepts (zyme accept) or rejects. Default budget: 50 rounds.

In a fresh coding-agent session from the task directory, paste:

272 lines · 31,368 chars
274 lines · 31,420 chars

Time: hours of unattended runtime. The agent commits per round, so you can interrupt and resume in a new session.

What to watch for:

  • results.tsv last keep row per dataset: the converged speedup
  • pipeline/run.{py,R} ends up holding the converged fast version
  • The agent runs the full budget (50 decision rounds) unless interrupted; it does not stop early just because recent proposals failed to improve

Stop here and run the Auditor again. Copy the optimization auditor prompt below and paste it into a fresh coding-agent session from this task directory. Continue to validation only after the verdict is PASS or WEAK.

147 lines · 16,350 chars

Optional: reflect before you move on. While this session is still open, run the Reflection agent right here (no fresh session) to report anything that tripped up the optimization loop. Recommended before you contribute.

101 lines · 6,237 chars

Next -> Auditor agent.