State-bound intent: solving the stale state problem in AI agent workflows

Short summary

The article introduces the 'stale state problem' in AI agent workflows: an agent reads a system state, reasons about a change, but by the time it submits the write, the target state has moved. The proposed solution is 'state-bound intent' — attaching a state reference (like a Git commit hash) to every proposed effect so the boundary layer can reject drifted requests before impact. This mirrors established patterns like Git base commits, database version checks, and HTTP ETags, but becomes critical because agents widen the read-to-write gap unpredictably.

• •Agents can reason correctly over stale state, producing requests that look valid but are based on a world that no longer exists
• •State-bound intent attaches a state reference to each proposed change, letting the boundary reject drift before impact
• •When state mismatches, the correct next action is re-reading the target and submitting a fresh intent — not retrying the same write

Generated with AI, which can make mistakes.