AI agent actions glossary: execution layer, bound approval, shared limit and receipt grades

This glossary defines the terms used when an AI agent stops recommending and starts acting: refunding, emailing, changing files. Each entry opens with a one-sentence definition. Each linked page adds an example and the limits of the term.

The terms describe one path. The agent names an action. Anlyon makes the call.

Without Anlyon   agent -> your API credential -> production API
With Anlyon      agent -> named action -> Anlyon -> production API

What sits between the agent and the API?

Execution layer for AI agents

An execution layer for AI agents receives the operation an agent requests, enforces the controls on that path, makes the external call and records its outcome. An authorization layer answers only whether the operation is allowed. Production execution for AI agents means the same thing applied to real systems: turning an agent's requested operation into a call against a real system, with controls and an outcome record. Execution layer vs authorization layer

Credential custody ("the agent never holds the key")

Credential custody means the provider credential for an AI agent's action, such as a Stripe or GitHub key, is held and used by the system that makes the call, so the agent names the action and never holds the key itself. It covers credentials in Anlyon's vault on the hosted path, not a key your own process already holds. Credential custody

Governed action

A governed action is an action an AI agent invokes through Anlyon that carries an adapter or a declaration, so it draws from a shared limit, its approval is bound to the exact request, and it returns a receipt with a grade. An action with neither keeps the template executor and gets none of the three. Governed action

What has to be true before an action runs?

Bound approval

A bound approval is a human approval tied to the exact request it was given for, so the request that runs is the request that was reviewed, and any difference is refused before dispatch. The human approves the exact change, and that exact change is what runs. Bound approval

Approval drift

Approval drift is any difference between the operation a reviewer approved and the operation that actually runs, such as a changed amount, recipient, target, action version or provider state between the decision and the call. A bound approval catches the request side. A provider's conditional write catches the provider side, where one exists. Approval drift

Shared limit (impact limit)

A shared limit, called an impact limit in Anlyon, is one allowance for what AI agents' actions change, such as money refunded or emails sent, that every agent, session and key in an environment draws from and that is reserved before each action is sent. A new session or key does not get a new allowance. It is shared within one environment, not across environments. Shared limit

Idempotency key for agent tool calls

An idempotency key for an AI agent tool call is an identifier for one intended operation, sent with every attempt of it, so a repeated call returns the recorded result instead of performing the side effect again. Two different keys are two operations, so a shared limit is what bounds the total. Idempotency key

What happens after an action is sent?

Action receipt and its grades

An action receipt is the record of what happened when an AI agent's action was sent to a provider, with a grade that says how the outcome was established: confirmed, acknowledged, unknown, failed, refused or denied. A receipt is a database record with provider evidence. It is not signed and not tamper-evident. Action receipt

• Confirmed: a read-back matched.
• Acknowledged: the provider accepted the request and nothing read it back.
• Unknown: no usable response, and Anlyon does not send the request again.
• Failed, refused and denied: it did not run. Failed means the provider rejected the request or it never left Anlyon. Refused means Anlyon stopped it at dispatch. Denied means the approval was denied.

Unknown outcome

An unknown outcome is the state of an AI agent action that may have reached the provider but returned no usable response, so nobody can yet say whether it happened, and it must be settled from the provider's own record before anything is retried. Anlyon records it as unknown, never failed, and does not send the request again. Unknown outcome

Halt (kill switch)

A halt, or kill switch, stops every AI agent in an environment from acting at the point where actions are sent, including calls already approved, while what you send keeps queuing so nothing in flight is lost. It stops new dispatch. A request an external server already accepted is not recalled. Halt

Where do these terms apply?

Every term above describes actions routed through Anlyon's hosted executor. A tool your own code calls directly is not routed through Anlyon and is not governed by it. The local approvals.gate() wrapper waits for an approval decision before calling your function, and your process runs that function with your credential.

Anlyon works on Stripe refunds, GitHub file updates, Resend email sends, and any HTTPS API you declare as an action. Stripe refunds run against your own Stripe account, with a live or a test key.

Start with a test action · Execution documentation · All guides