Most agent accountability efforts focus on identity — proving who an agent is or who authorized it. A new open-source project called COGEXT, described in a dev.to post this week, tackles a narrower and more mundane problem: agents constantly say they'll do things, and nothing tracks whether they actually do. The author's framing is blunt. Agents are good at generating intent and bad at remembering it — they'll promise to email a client, confirm an appointment, or deploy a fix, and forget they said anything a couple of turns later. The insight driving the project is that this isn't a memory problem to solve with bigger context windows; it's a state-machine problem. COGEXT sits between an agent and the outside world. It ingests agent output, extracts every commitment made in it, and tracks each one through a 12-state lifecycle — DETECTED, OPEN, DUE, OVERDUE, FULFILLED, FAILED, EXPIRED, PENDING_REVIEW, CANCELLED, CONTRADICTED, BLOCKED, and back to OPEN or FAILED depending on how a blocked commitment resolves. The core distinction the author draws is between logging and tracking: most systems treat agent output as text to archive, while COGEXT treats it as events with a lifecycle and an expected outcome. The implementation detail worth noting is the atomicity requirement. State transitions and audit events are written in a single database transaction — a status change to "fulfilled" without a corresponding fulfillment event in the log is treated as a bug, not an edge case. That's a small design choice, but it's the same principle underlying every serious attestation scheme this year: a claim about what happened is only as trustworthy as the evidence trail proving it happened, generated at the same moment as the change itself, not reconstructed afterward. What's notable here isn't novelty of the state machine itself — commitment tracking with due dates and failure states is a well-worn pattern in workflow software. It's the application to agent output specifically, as a distinct accountability layer sitting above whatever protocol the agent uses to talk to tools or other agents. As multi-agent systems increasingly delegate open-ended tasks across turns and sessions, "did the agent do what it said it would" is a surprisingly under-built question. Commitment tracking won't tell you whether an agent was authorized to make a promise, but it will tell you, unambiguously, whether it kept one.