How are you handling guardrails around the self-rewrite step?
Specifically curious whether revisions are gated by deterministic checks or learned evaluators, and how you avoid runaway feedback loops.
We wanted to explore what agents look like when you design for production instead:
– memory pressure
– cost ceilings
– failure recovery
– observability
So we built an open-source framework where agents can:
• refactor their own logic
• revise plans mid-execution
• persist and compress memory
• recover from tool and runtime failures
This is not a wrapper around prompts — it’s infrastructure for long-running, self-correcting agent systems.
We’re building this in public and actively using GitHub as our distribution channel.
Feedback from people deploying agents in the wild is especially welcome.
If it’s useful, starring the repo helps us decide what to build next.
How are you handling guardrails around the self-rewrite step? Specifically curious whether revisions are gated by deterministic checks or learned evaluators, and how you avoid runaway feedback loops.