AI Jobs

An AI job is a standing order for an agent: what to do, when to do it, and where to keep what it finds. You describe the work once. From then on it runs by itself on a schedule, or whenever someone presses Run now, and every run ends with a report.

Jobs are for work that comes back:

Example What the job does each run
Watch a list of sites for announcements Opens every source, records what it has not seen before, keeps a list of what is current
A weekly digest Reads what happened in the workspace or on the web and writes the summary
Keep a summary fresh Rewrites one living document every morning from its sources
Check something on a schedule Runs the same check and reports only what changed

A job is not a conversation. It runs in the background with no one watching, so its instructions have to stand on their own, and what it learns has to be written down so the next run can build on it. That is what artifacts are for.

Where jobs live

AI Studio → Jobs in the sidebar. AI Studio also holds Agents (the agents that run jobs), Workflows and Runs (everything agents are doing for jobs, in one place).

The parts of a job

Part What it is
Agent Who runs it. Any agent in the workspace: a server-side agent for light work, a harness agent when the job needs a computer of its own — files, code, a browser, PDFs.
Brief What the job is for and the facts it always needs: the sources, what to record, how to name things. Read at the start of every run.
Prompt or workflow What to do on each run: a prompt in plain words, or one of the workspace's AI Workflows with its arguments.
Trigger When it runs: by hand, on a schedule, or on a cron expression.
Artifacts What the job keeps between runs: living documents rewritten every run, and items that accumulate.
Subscribers Who hears when a run ends: in the Inbox and by email.

How a run goes

  1. It starts at its scheduled time, or when someone presses Run now.
  2. The agent reads the brief, the prompt and the job's artifacts — everything earlier runs left.
  3. It does the work, saving and archiving artifacts as it goes.
  4. It writes a report: what is new, what changed, what it could not do.
  5. Subscribers are told the run finished or failed.

A job has one run at a time. Run now is unavailable while a run is queued or going.

In this section

  • Creating a Job — the form: agent, brief, prompt or workflow, trigger, subscribers.
  • Runs and Reports — the jobs list, a job's page, the Runs page, cancelling, notifications.
  • Artifacts — what a job remembers: living documents, items, revisions, archive, reset.
  • Writing a Job That Works — how to word a brief and a prompt so runs are complete, consistent and cheap.

Agents can work with jobs too. An assistant connected over MCP, or a server-side agent in a discussion, can create, change and run jobs and read their runs and artifacts, with the rights of the person it works for.