← back
Your Agent Didn't Fail. Your Harness Did. — Vinoth Govindarajan, OpenAI
Takeaway
Production agent reliability depends on durable state ownership, ordered mutations, and verifiable action receipts in the harness.
Summary
- An agent can deliver a successful-looking reply while failing to persist the underlying event, leaving later turns confidently reasoning over incomplete history.
- The harness must own state transitions, authority checks, ordered commits, and receipts that distinguish proposed actions from executed and confirmed outcomes.
- Each durable fact needs a named system of record and replay path; shared mutable state needs a single ordered commit path while independent reads and work can remain parallel.
- Dangling tool calls require deadlines, cancellation, watchdogs, explicit terminal results, and recovery commands that bypass blocked work.
agent-harnessdistributed-systemsreliability
Original description
Two runs touch the same session, the second write silently erases the first, and the agent keeps answering with total confidence from stale state. Nothing crashed and the model did not hallucinate, so this is a harness failure, the kind that lives in the system around the model rather than in the weights. Using OpenClaw as a public case study, Vinoth Govindarajan walks the usual suspects: state that was never persisted, overlapping writers with no single writer lane, a tool call that never returns because nothing set a deadline, and an approval that outlived the action it was supposed to authorize. The through line is that a model only proposes; the harness has to commit, and a receipt has to prove it. A transcript shows what the agent said, but a receipt is the evidence that survives: it records the mutation, the authority used, and whether the message actually reached the user, since an internal success that never becomes visible proof is its own failure. You leave with a run receipt audit to run on your own agents, five questions per incident: what woke it up, what state did it inherit, what authority did it use, what executed, and what evidence survived. Speaker info: https://x.com/iamvinoth / vinothgovindarajan https://theagentstack.substack.com/ Timestamps: 0:00 - Introduction: harness failures vs model failures 1:32 - Delivery can succeed while the truth fails 2:46 - A model proposes, the harness commits, the receipt proves 4:14 - How events enter and state is rehydrated 5:48 - Idempotency, locks, and ordering 7:22 - Ownership: who persists the turn 8:28 - Single writer lanes and overlapping writes 10:09 - Time, deadlines, and cancellation 11:23 - Approval drift and bounded authority 13:05 - Internal success vs user-visible proof 14:08 - The run receipt audit: five questions