Handing real work to agents raises two questions, at two different moments.

Before the work: what is this agent allowed to do? After the work: was the result right, and who said so?

Most teams answer the first with permissions and the second with a thumbs-up in chat. OrgX keeps both answers on one record, attached to the same initiative.

Before: guardrails

Every initiative in OrgX runs in one of three autonomy modes.

• Manual. Every step waits for a person.
• Gated. The default. Low-risk steps continue on their own. A deliverable that crosses a boundary, like an opened pull request or a submitted draft, waits for review unless you granted that kind of work ahead of time.
• Autopilot. Low-risk steps and granted deliverables continue.

Under all three sits a floor that no mode lifts: sending, spending, deploying, merging, and deleting always stop for a person. An agent cannot raise its own autonomy, and it cannot approve its own work. Both rules are enforced in code and fail closed.

Each connector carries its own setting too: allow, ask, or deny.

After: acceptance

Guardrails keep an agent inside the lines. Whether the work inside the lines was any good is a separate question. For that, each piece of agent work carries:

• The claim. What the agent reported as finished.
• The decision behind it. Decisions keep their provenance across sessions and clients, so the next agent inherits the reasoning instead of a summary.
• The check. The published quality rule the work was evaluated against, and whether it held or cleared the work. This part is in beta.
• The ruling. The owner who accepted the work, and when.

Why both live in one place

Our own ledger shows what happens when only one side gets built. In July we published the number: 3,328 execution receipts and one acceptance row. Recording what agents did scales on its own. Recording what a person accepted only happens when the system gives that ruling a place to land.

When the limit and the ruling sit on the same initiative, you can answer the question a client, a manager, or an auditor actually asks: what was this agent allowed to do, what did it do, and who signed off?

Try it on one piece of work

Bring one recent piece of agent work to a 30-minute proof review. We will walk through what the agent was allowed to do, what it claimed, and what a published rule would have held. Or start free with one initiative and read the Agent Work Receipt v0.1 spec.