DocsBuild an agent
Watching for changes
Be told when an answer changes: a watched stack, the answer change emails for models, and the Atom feeds of the log.
An answer is a reading of the catalogs on one day, and the catalogs are re-read every day. Three things tell you when a later reading differs: a watched stack, the answer change emails for models, and the log's feeds, which need no account.
A stack
An answer is true on the day it is given. A server's owner publishes a new version, changes its tools, or stops answering; an advisory is listed against the version in your config. To be told, save the stack: on Architect an agent, sign in and press Watch under an answer. The need is resolved again each day, and you get one email on a day something changed: a different choice, a config that names another version or asks for something else, or an entry in the log about a server or skill in the stack. For a stack that runs on something, with the vendors you named, it is also a day the least of them is another product, what its reference job comes to moved, or the log recorded a price of one of them. On a day nothing changed there is no email.
What is stored is the structured need (the systems by id, where the agent runs, the harness, what it does and needs, the infrastructure products you named by id) and the name you give it. A month of usage is not kept: what a stack runs on is watched as a reference job, at listed rates. A watched stack asks for no model; a model's changes are the answer change emails' part, below. Your watched stacks are in the console, where one can be paused, renamed or dropped, and each email has a link that pauses them with no sign-in. A plan watches a number of stacks: 1 on Free, 5 on Builder, 25 on Pro and unlimited on Scale. An account whose plan becomes smaller keeps the stacks it watches, and their emails; it cannot add another until it watches fewer than its plan's number.
There is no API call that sets a watch. An answer carries watch_url: the form with the need filled in and the watch panel asked for. It is left out when no system was recognised, and for a need with prefer or a tenancy of multi, which the form does not have. A coding agent passes it on to its user, who opens it and presses the button; opening the address saves nothing.
A model answer
A model answer changes only when the model catalog does. To hear about it without polling, an admin or the owner turns on Notify me when my answers change in Settings (off by default). While it is on, each successful /v1/resolve and /v1/plan call made with a live key keeps the need it resolved and the answer: the task or purpose, providers, constraints and shape, the key's project, and the chosen model, provider and provider model id of each member. For /v1/plan it is only what the request stated, with the defaults: never the prompt, tags or attachments, nor a value read from the prompt or attachments, such as its size. A need naming an unknown id or role is not kept. Test keys and the sandbox are never recorded.
After each daily catalog refresh, every kept need is resolved again. When an answer changed, no longer resolves, or resolves to a model that retires within 30 days, the owner and admins (verified emails) get one email that day, for example: "The answer for face-bananas' photo-edit-with-reference / gemini changed from … to …", or "… retires on 2026-12-15". Up to 100 needs are kept, the least recently used dropped first; a need unused for 30 days is deleted. Turning the setting off deletes them all at once, and the change shows in the audit log.
The log's feeds
The log records what differs between each published agent catalog and the one before it (see how it is made). To follow it without an account, the whole log is an Atom feed at /feeds/today.atom, and each system has one of what the log records about it, at /feeds/systems/<system>.atom. An entry's id never changes, and a feed changes only when the catalog does.