An agent can finish a computation and leave the actual question unresolved. The tool call succeeds, the transcript looks clean, and the thing you cared about is still open. Most systems treat that gap as a prompt problem. I think it is a protocol problem.

Closer Protocol is my proposal for what should survive between agent episodes: the obligation. Not the transcript, not the plan. The obligation. A retained question with an origin, a scope, resolution conditions, and an authority boundary. New evidence connects to it or it does not. Proposed action against it gets checked against what the user actually authorized, not what the model inferred from a broad goal. It stays open until an adjudicated commit says it is closed, and while it waits, the system works on something else instead of pretending.

None of the parts are new, and the document says so up front. Actor systems, durable execution, event sourcing, truth maintenance, reference monitors, human approval workflows. All established, all cited. What I am claiming is the combination, and only as a hypothesis. Whether this composition beats a simpler, equally resourced workflow is an open question with a test design attached, not a result. Section 10 is the part I intend to hold myself to.

The document was drafted by AI agents under my direction, and it accounts for exactly who contributed what. That accounting is not a disclaimer. It is the point. A protocol about evidence-governed work should be able to show its own receipts.

I am publishing it here so it has a stable address. I want to reference it from anywhere, test the hypotheses against real work, and let anyone who reads it tell me where it breaks. It is a review draft. It will change.

Closer Protocol v4 (PDF) · Markdown source