# Your Agent Didn't Fail. Your Harness Did. — Vinoth Govindarajan, OpenAI

## Executive summary

The talk argues that most production failures in AI agents are not due to model hallucination or poor reasoning, but rather 'harness failures'—flaws in the underlying system architecture. The core principle is establishing a clear boundary: while a model can only propose an action, the surrounding harness must own the state transition, enforce ordered commits, and provide verifiable proof (the receipt) that the action was successfully executed and observed by the user.

## Key takeaways

- Model Proposes, Harness Commits, Receipt Proves: The model generates capability; the harness provides control. The system must own state transitions, enforce ordered mutations, and use a receipt to prove that an action was executed and visible to the user.
- Own the State, Order the Mutation, Prove the Action: These three principles are critical for reliable agent systems. A fact must have a single owner (system of record) and a clear replay path; mutable state requires one ordered commit path.
- The Run Receipt Audit: To audit an agent system, ask five questions: What woke it up (trigger)? What state did it inherit? Which authority did it use? What executed? And what evidence survived?

## Technical details

- State Ownership and Persistence: The system of record (owner) must be defined for every fact. If no named owner can reliably replay a fact, the agent has not truly remembered it. The harness rebuilds working state from multiple sources (transcript, session state, memory, policy).
- Overlapping Writes and Concurrency: The 'last writer wins' approach is insufficient. To maintain consistency, the system must enforce one ordered commit path for a mutable state boundary using mechanisms like queues (Q), mutexes (mutx), transactions, or locks.
- Lifecycle Failure Modes: The run cannot wait indefinitely. Every external boundary needs a terminal outcome: Success, failure, timeout, cancel, or max attempts. Runs must implement deadlines and cancellation mechanisms to prevent the system from getting stuck.
- Authority Scoping (Approval Drift): Capability is not execution; requestability is not authority. Approvals must be treated as scoped execution states, binding to a specific action and remaining tied to it even during retries or callbacks. The principle of least privilege should narrow the tool surface.
- Proof vs. Claim (Missing Edge Proof): Internal component success does not equate to external proof. A successful API call only proves the internal path accepted the request; it does not prove the user saw the result. The receipt must confirm the outcome at the user-visible boundary.

## Practical implications

- Implement a formal 'run receipt' mechanism to record the mutation, authority used, and final user-visible outcome for every agent turn.
- Design state management systems with explicit ownership boundaries (system of record) rather than relying on implicit memory or transcript history.
- Enforce ordered commit paths using concurrency controls (locks/transactions) when multiple sub-agents can modify the same mutable state.

## Topics

AI Agents, Distributed Systems, State Management, Concurrency Control, System Reliability Engineering, OpenClaw, The Agent Stack

Source: https://www.youtube.com/watch?v=BInpv7lGp1o