Auto-Research

Goal

• Run bounded autonomous optimization loops with measurable metrics.
• Keep experiment state on disk so a fresh agent can resume exactly.
• Prefer simple changes that survive verification, not random mutation spam.

Use this when

• The user wants to optimize a measurable target: test speed, build time, bundle size, training loss, memory, prompt quality, workflow latency, etc.
• There is a clear benchmark command or evaluation harness.
• The work has a rollback path (git, branch, snapshot, or disposable environment).
• You want autoresearch-style keep/discard loops inside Hermes.

Do not use this when

• No primary metric exists.
• The benchmark cannot be run repeatedly.
• The task is destructive/high-risk and lacks approval.
• Success is mostly subjective and cannot be reduced to stable checks.