Lobster Pot
Built at WeaveHacks 3: Self-Improving Agents Hackathon with Weights & Biases · Jan 31, 2026 · San Francisco, CA
Lobster Pot is a safety-first agent guardrail system for OpenClaw that forces all actions into a sandboxed environment "lobster trap" before execution, logs everything through Weave for auditing and evaluation "lobster pot", and uses recursive prompt refinement "keep cooking" to prevent unsafe or wasteful agent behavior over time. By analyzing traces with RLM-style long-context reasoning and iteratively patching prompts/policies within the session, Lobster Pot stops cost blowouts, prompt drift, and exposure risk while learning how you want tasks run. Only the hardened shells get released into the sea! Try it out: https://weavehack3-tsx7-6n7je5ot5-ryans-projects-a4a34d15.vercel.app/ https://youtu.be/Z25eVeUDY2E for better concept page