The Work agent: manage it like a hire.
Our free ChatGPT Work lesson covers what the agent is. This is the playbook for using it seriously: the scoping discipline, the supervision rhythm, and the review habits that separate people who get finished work from people who get expensive drafts.
OpenAI launched “dots” — agents that keep working between conversations. Per OpenAI, a dot is an always-on agent powered by GPT-6 Astra with its own cloud computer and browser, connected to apps through plugins (OpenAI cites over 4,000) and reachable from ChatGPT, Slack and Teams. It is rolling out to Pro and Business Premium users in eligible markets (Enterprise in beta when an admin enables it); the release notes say Pro access excludes the EEA, Switzerland and the UK at launch and that users must be 18+. Your first dot is included in the plan, with an allowance for deeper work. OpenAI lists read-only background access, auto-review of sensitive actions and custom permission rules as built-in controls. Everything in this playbook applies with extra force to something that runs unattended: scope it tightly, start read-only, and review its activity log.
01 The mental model: a talented contractor, day one
ChatGPT Work will happily accept a vague brief and burn hours producing the wrong thing beautifully — exactly like a talented contractor with no context. Every skill in this lesson is a management skill: brief well, check in at the right moments, review like it's your name on it. Because it is.
02 Scoping: the brief format that works
Deliverable: exactly what comes back ("a 10-slide deck," "a cleaned sheet plus summary memo"). Inputs: which files/plugins to use — name them. Definition of done: the checklist it must satisfy. Constraints: what it must NOT do (no emailing anyone, don't touch the master file, cite every number). Taste reference: attach one example of what good looks like. The example does more work than any adjective.
Scope one deliverable per job. "Research competitors AND build the deck AND draft the email" is three jobs; run them as three, feeding each result into the next. Chained small jobs beat mega-jobs on quality and on usage cost.
03 Supervision: the check-in pattern
04 Review: agent failure modes to hunt
- The confident gap: it couldn't find something, so it produced something plausible instead. Check every number against the named source.
- Instruction drift: hour-three output quietly forgets hour-one constraints. Re-read your "definition of done" against the result, item by item.
- The over-polish: beautiful formatting wrapped around thin substance. Grade content first, with formatting mentally stripped.
- Scope creep it invented: extra sections nobody asked for. Fine to keep, but recognize it — it's usage you paid for.
05 Usage budgeting
Work consumes plan usage in proportion to task size — a multi-hour job costs real capacity (it meters like Codex, not like chat). Budget like a manager: run your first week's jobs small deliberately to learn the burn rate, keep a rough log of job-size versus usage, and save the mega-jobs for where finished-work value is obvious. On desktop, remember it can also use local files and apps — same rules, plus the connector security thinking applies double.
Write the full five-part brief for a real deliverable you owe someone this week. Run it. Do the 5-minute plan check. Review against your own definition-of-done list.
Open ChatGPT →This week's challenge
Run three scoped agent jobs this week, chained: research job feeds a build job feeds a draft job. Log brief quality against output quality. By job three you'll have learned the thing every new manager learns: the output was determined by the brief.