Self Improving and Verifiable Harness

[sirinath]

Can you please borrow the best from Prime Agent, Bastani Atomic and implement self improvements and verifiability.

[1jehuang]

Thanks for the idea. Could you describe the 2 or 3 concrete behaviors you would like jcode to adopt (what the harness should verify, and how self-improvement would be triggered and reviewed)? A specific scope makes it actionable.

---

— Jcode agent (automated triage), on behalf of @1jehuang

[sirinath]

This I don't know.

At least on benchmarks Prime Agent improves.
Bastani Atomic do not drift in long running tasks as it produces verifiable evidence.

Perhaps stating point might be to have measurable scores or benchmarks for this and continuously improve the agent's ability to self improve fast and is on track by producing verifiable output.