Release is the act by which a named person lets an action an AI agent has prepared and staged take effect: a payment file submitted, a trade sent to market, a letter posted. The agent does the work up to the point of no return; the person decides whether to cross it. A release is a decision, recorded with who made it, what exactly they released and when.
Why does release matter once agents act?
Because it is where accountability is placed. When an agent prepares a payment run and a finance manager releases it, the manager owns the decision to pay. That only works if the release is real: the person saw the actual action, had the authority to approve it, and could have refused.
Many approval steps fail this test. A person clicks "approve" on a summary, the underlying file changes afterwards, or the approver lacks authority for the amount. The record shows a release that did not control anything. That is why we hold that agents prepare payments and people release them, with the release bound tightly to what it covers.
How does a release work in practice?
The agent stages the complete action, with the facts it rests on and their provenance. The platform holds it and presents it to a person entitled to release it. That person reviews and releases, edits, or refuses. Four checks make the release meaningful:
- Exact object. The release applies to the staged action as presented. If anything changes after release, it needs a new one.
- Right person. The releaser holds authority for this action and amount, and is not the person or agent who requested it.
- Full view. The releaser sees the action itself and the facts behind it, not just a summary line.
- Record. Who released, what, when, and on what evidence is filed with the verdict.
For bank payments the steps are set out in how to prepare bank payment files for release with an AI agent; for trades, in how to let an AI agent stage trades for a person to release.
What does that look like in an example?
A charity's finance agent prepares the monthly grants payment file: 40 grantees, amounts from the approved grants register. The finance director reviews the file in a single view showing each payee, amount and the register entry behind it. She removes one payment whose bank details were changed by email a few days earlier and releases the rest. The release records the exact file she approved. When the agent re-stages the removed payment after a call-back confirms the new details, it comes back to her as a separate release.
How is this release different from a software release?
In software, a release is a new version of a system put into use, sometimes gradually through a canary release. In agent governance, release is a person allowing one staged action to take effect. Both control what reaches production, but one governs code and the other governs individual business actions. An agent's new brief or model goes through a software-style release; each payment it prepares goes through an action release.
How does release relate to the autonomy ceiling?
Actions above an agent's depth, or above its autonomy ceiling, are held for release by default. As an agent earns depth for a capability, some actions that once needed release can proceed without it within bounds. Some never move: certain actions are reserved for a person whatever the evidence. The reasoning is in regulated autonomy.
Questions people ask
Can an agent release its own staged action?
No. A release by the same agent that staged the action is no control at all. The releaser must be a person with authority for that action, separate from the requester.
Can release be done in bulk?
Yes, for batches of similar items, provided the person sees each item and the release records exactly which items it covers. A single click over an unseen list is not a release.