Durable Agent Engine
Fault-Tolerant LLM Agent Orchestrator
Screenshots
Overview
Most agent demos show the happy path. This one shows what happens when a worker crashes mid-step, two workers race for the same job, or a tool call fails — and proves the run still finishes correctly, exactly once, every time.
Technical Details
Built a Postgres-native durable queue for multi-step LLM agent DAGs — no Redis, no queue library — with SKIP LOCKED exactly-once claiming, a lock-TTL reaper that reclaims work from crashed workers, and exponential-backoff retries into a dead-letter table.
Proved crash-safety with a chaos test injecting 40% failures plus kill -9 mid-run: every run still reaches a terminal state, with an append-only audit log of every state transition.
Live demo streams each step's claim, retry, and completion over WebSockets so the failure handling is observable rather than asserted.