Governing AI agent work

Handing a ticket to an AI agent is easy. Knowing what it did — and stopping two agents from doing it at once — is the hard part. AgentTask treats governance as table stakes, not an enterprise add-on.

Claims: one worker per ticket

When an agent starts a task it takes a claim — a lease with a fencing token that other workers must respect. A second agent asking for the same ticket is told who holds it instead of silently duplicating the work. Claims heartbeat while work is live and expire if the worker dies, so stuck tickets free themselves.

Attribution: every write has an actor

Writes made with an OAuth token are attributed to that user, and writes made with an organization API key to the key’s creator; a credential with no accountable user is refused. Fleet workers act under a claim, so their writes are recorded against the claim and the agent that holds it. When an agent posts an update, you can always answer who ran it, with whose access, on which ticket.

The audit trail is the product

Plans, runs, tool calls, and deliveries are tied back to the task, and requests made to the API and MCP land on an append-only audit trail (kept for a limited retention window, so export what you need to keep). That turns “what did the agents do last week?” from an investigation into a query — and it is why teams can let agents touch production repos without holding their breath.

If your team is evaluating how to run agents on real work, this is the shape we’d argue for: shared backlog, claimed work, attributed actions, and a request-level audit trail.