Agent intent

What it is

Most logs say what happened: this key called this action at this time. Intent adds why, in the agent's own words, at the moment it acted: "The customer asked by email this morning to cancel their backup add-on." It is a short sentence, written by the agent, attached to the call and stored with the record of it.

Why it matters

• Review. A person reading an audit log can tell a sensible action from a strange one without reconstructing the whole conversation.
• Spotting manipulation. An action whose stated reason does not match the task is a signal, for example a "routine sync" that exports every contact. See prompt injection.
• Approvals. A person deciding whether to let an action run needs the reason as much as the arguments. See human-in-the-loop approval.
• Accountability. Logs that record only identities and endpoints make investigation slow. OWASP's agent cheat sheet asks for agent decisions, tool calls and outcomes to be logged, and for high-risk actions to carry structured context such as the approval and the policy that applied.

How it works

1. The tool or API makes a reason a required field on every action, not an optional note.
2. The agent writes one plain sentence in its own words, specific to this call.
3. The platform stores it on the same record as the action, its inputs' hash, outcome and cost.
4. Grouped work, such as one question sent to several systems, shares one intent across every call it makes.
5. A retention rule decides how long the free text is kept, because a reason can quote personal data.

Example

do service.ticket.create · intent: "Logging the backup failure the client reported by phone at 9:05 so the on-call engineer sees it."

Six weeks later an auditor asks why that ticket exists. The answer is on the record, next to who created it and when.

Common pitfalls

• Boilerplate. "Performing task" on every call tells nobody anything. Ask for the specific reason.
• Treating intent as proof. It is the agent's claim. A mistaken or manipulated agent can state a plausible reason; permissions and approvals still have to hold.
• Personal data in free text. Reasons quote customers. Keep the text for a defined period, not indefinitely.
• Optional fields. If the reason can be left out, it will be, on exactly the calls you later need to explain.

How actionlake handles it

Every do and ask call must include an intent: one plain sentence, in the agent's own words, saying why it is making the call. A call without one is refused. The intent is stored on the run's row in the ledger; a read sent to several systems at once is one grouped run, and every call under it carries the same intent, as do the runs of a plan. The intent text is cleared after 90 days; the record of the action, with its hashes, ids, cost and outcome, stays. See the trust page.