Author a policy through your agent

Author an Agent Intent Policy (AIP) by working with your own agent, from a brief to a ratified policy, without leaving your agent. The agent does the legwork: it reads your brief, runs the interview and prepares the checks. You decide, and nothing takes effect until you approve the ratify call.

You need an agent with the OpenLatch plugin installed, which brings the MCP connection and the openlatch-autonomy-zones skill, and the editor role or higher. Write an Agent Intent Policy describes the same steps in the console, and Intent policies explains the model.

From a brief to a ratified policy

Ask your agent to write a policy, and tell it the zone the policy is for. The agent follows the skill’s steps with these tools:

Step	What the agent does	What you do
Start	Reads the limits with drafts_read, then creates a draft from text you paste with drafts_write, or from a built-in preset	Give the brief in your own words, or choose a preset
Interview	Shows each intent policy’s interview cards, with the recommended answer, and sends your answers with answer	Answer every card; the agent never answers one for you
Confirm	Asks Is that right? for each sentence and confirms it with patch_policy only after you say yes	Keep a sentence, change it, or leave it out
Zone and mode	Sets the zone and the mode with set_zones	Choose Monitor or Enforce; start in Monitor unless you already trust the policy
Prepare	Prepares the draft with prepare, polls the job, then reads the test cases and the replay with drafts_read	Read the failing cases and the replay
Ratify	Prints every sentence the call puts in force, then calls policies_ratify	Approve the call in your client

A write that starts a job returns a job the agent polls with job; it never waits inside a call. The agent can start only from text and from presets. To upload a document or record a conversation as the brief, use the console.

Approve the sentences

Ratifying sends the exact sentences it puts in force, in the sentences argument of policies_ratify. Your client shows that argument in its permission prompt, so you read what goes into force before you approve. The platform then checks the sentences against the draft inside the ratify itself:

What the agent sends	Result
The draft’s sentences, copied word for word	The policy is ratified in its zone
No sentences	The tool refuses the call before it reaches OpenLatch; nothing changes
Any other sentences: a paraphrase, one missing, one from an earlier revision	Refused with SENTENCES_MISMATCH; nothing changes

The comparison ignores the order of the sentences, extra whitespace and the way Unicode composes a character, and nothing else: a changed word is a mismatch. The refusal never repeats a sentence back.

OpenLatch marks policies_ratify as needing your interaction, and some clients shorten long arguments in the prompt, so the skill has the agent print every sentence, with its dimension, the zone and the mode, in its own message before the call.

If you set your client to allow policies_ratify without asking, there is no prompt. The check then proves only that the agent sent the draft’s current sentences, not that you read them. Keep the prompt on for this tool.

Edit a policy or restore a version

To	The agent does
Edit a ratified policy	Reads the policy with policies_read, then opens an edit draft with edit. The live policy keeps enforcing unchanged until you ratify the draft as a new version. One person edits a policy at a time, and the agent tells you who holds it
Restore an earlier version	Reads that version’s sentences with the version method of policies_read, shows them to you, then calls policies_ratify with restore and those sentences. The restore makes a new version and rewrites no history

Both go through the same sentences check as a new policy.

Delete a policy or copy it to another zone

A policy keeps its Autonomy Zone for its whole life, so the agent never moves one.

To	The agent does
Delete a policy	Calls policies_write with retire. After your approval the policy stops running and cannot be used again, and its history stays in Audit
Run a policy’s rules in another zone	Calls policies_write with copy and the target zone. It creates a draft in that zone, in Monitor, which you review and ratify as a new policy with its own history

A call that names attach or detach is refused with OPERATION_REMOVED, and the agent is told to delete or copy instead. Nothing changes.

When a call is refused

Code	What happened	What the agent does
SENTENCES_REQUIRED	The ratify carried no sentences	Reads the draft, shows you its sentences and sends them
SENTENCES_MISMATCH	The sentences are not the ones this call puts in force	Reads the draft again, shows you what it holds now and sends exactly those
DRAFT_NOT_PREPARED	The draft was not prepared at this revision	Prepares it again, polls the job, then ratifies
DRAFT_REVISION_STALE	The draft changed after the agent read it	Reads it again and shows you what changed
EDIT_DRAFT_STALE	The ratified policy changed after the edit draft opened	Reads the policy again and opens a new edit draft if you still want the change
DRAFT_EXISTS	Someone else holds an edit draft of the policy	Tells you who holds it
AI_NOT_ENABLED	AI authoring is off for your organization	Starts from a built-in preset instead
FORBIDDEN	Your role does not allow authoring	Tells you which role it needs

Every code is listed with its status in the error catalogue.

Next, graduate the policy to Enforce once its record shows it is accurate.