OpenAI’s dots turn an ongoing responsibility into something you can delegate to an agent with its own computer. The useful first step is to choose a small responsibility with an observable result. Before connecting valuable accounts, decide what the dot may read, what it may change and which decisions must return to you.
Written October 1, 2026 and dated September 29, 2026, using the dated announcements cited below. Later product developments are outside this article’s scope.
- 01Access is stagedCheck your plan, market and workspace settings before planning a rollout.
- 02Reading and acting differProactive research is read-only; delegated work needs its own action boundaries.
- 03Included does not mean unlimitedTasks started in Work or Codex use their usual plan allowances.
- 04Review the exceptions, not the highlightsA finished assignment where a source was missing or an approval was required shows more than an impressive first answer.
01 — The evidenceWhat OpenAI announced
In its September 29 introduction, OpenAI describes GPT-6 Astra agents with their own cloud computer and browser, connections to more than 4,000 apps, and access through ChatGPT, Slack and Teams. Text messaging is planned. Pro and Business Premium access is rolling out in eligible markets; Enterprise, Edu and Healthcare workspaces must enable the beta. Specialist dots begin with enterprise pilots, while Microsoft Agent 365 integration remains in development.
Those are product capabilities and announced stages, not a promise that an agent can finish any job correctly. A persistent agent changes how work arrives and continues, but the person delegating still needs to define success. A weekly research brief, for example, can have a fixed source list, a delivery format and a reviewer. “Help grow the business” leaves both the actions and the acceptance standard undefined.
Our DevDay preparation guide covers the planning background. Dots add a continuing relationship to that picture: instructions need to remain understandable after the conversation that created them is no longer fresh.
02 — Practical implicationsWrite the responsibility as a short contract
Give the first assignment an input, output, cadence and stopping condition. An example is: review these public release pages each weekday, save a short brief with source links, and flag changes that affect our current product. Sending the brief to a client should be a separate decision. This makes the difference between preparing information and representing the business explicit.
| Part | Example boundary |
|---|---|
| Inputs | Only the named public pages and an approved internal product brief. |
| Output | A dated draft with source links and an uncertainty section. |
| Actions | Save the draft; request approval before external communication. |
| Stop | Pause if a source requires new access or the task exceeds its budget. |
Describe exceptions in plain language. If a source is unavailable, the agent should record that fact instead of substituting an unrelated source. If two sources disagree, it should show the conflict. If a task requires a login or a broader permission, it should return to the owner. The value of these instructions is that a reviewer can check them against actual work.
03 — ApprovalsSeparate background reading from authorized action
OpenAI says proactive research uses read-only tools. Custom Rules can allow, require approval for or block actions, and Activity View exposes progress. Auto-review checks consequential actions against instructions and safety requirements; some sensitive tasks stay with the user. These mechanisms reduce the amount of supervision needed, but they do not remove the need to review consequential outputs.
Collect a bounded evidence set
Use approved sources and inspect the resulting brief for omissions or unsupported claims.
Create a reviewable draft
Let the agent prepare a message, patch or document while you decide whether it is ready.
Change a business system
Authorize a narrow class of actions only after the preparation workflow behaves reliably.
Test a rule with an innocuous example before relying on it. Ask the dot to prepare an external message, then inspect whether it waits at the intended boundary. Check the activity record for the attempted action and the approval request. A rule that sounds precise can still leave ambiguity about recipients, attachments or which account is being used.
Our guide to permission defaults explains why narrow access matters independently of the quality of the instructions. Connecting an app creates a capability; the task contract explains when that capability should be used.
04 — BudgetingBudget the work, including follow-on tasks
OpenAI includes the first dot with Pro or Business Premium. Dot conversations do not consume ChatGPT limits, but tasks it starts or manages in Codex or Work count normally. The launch post does not publish additional-dot prices. Do not translate “included” into an unlimited monthly production budget.
For the pilot, record the number of requested tasks, accepted outputs, follow-on jobs and minutes of human correction. Those observations are more useful than counting messages. An agent can exchange many short messages while doing little billable work, or start a substantial task from one sentence. Put an owner on the allowance and define what should happen when it is nearly exhausted.
Review access and data settings for the specific workspace. Keep the first assignment away from unnecessary personal or customer information. A useful public-source brief can establish the workflow before the organization makes a separate decision about private material. Our runtime boundary comparison provides a broader checklist for computer access.
05 — The decisionExpand after reviewing complete outcomes
Review several finished assignments, including one where information was missing and one where an approval was required. Look for useful evidence, correct routing of exceptions and a clear record of what happened. An impressive first answer is less informative than consistent behavior across these ordinary difficulties.
If the pilot succeeds, expand one dimension at a time: more sources, a longer task or a new action permission. This makes a regression easier to explain and reverse. Our AI transformation service helps teams define these task-level evaluations and operating boundaries.
06 — The failuresFour first assignments that went wrong
A persistent agent fails differently from a chat: the mistake repeats on a cadence, in the background, until someone looks. The illustrative first assignments below each skip one section of this guide. They are design examples built on the research-brief contract above, not observed deployments.
The open-ended responsibility
The dot is told to help grow the business. With no input list, output format, cadence or stopping condition, both the actions and the acceptance standard are undefined, and the first week’s work cannot be checked against anything.
The substituted source
A named public page is unavailable one morning. The contract did not say what to do, so the brief is assembled from an unrelated source and reads as complete. The exception rule, record the gap instead of substituting, was never written.
The untested rule
A Custom Rule requiring approval before any external message sounds precise. Nobody asked the dot to prepare an innocuous message first, so the first real one goes to a recipient the rule did not cover, from an account nobody checked.
The included that was unlimited
The first dot is included with the plan, so nobody puts an owner on the allowance. Tasks it starts in Codex and Work count against their usual limits, and the allowance is spent on follow-on jobs nobody counted.
Each failure is visible in the activity record, after the fact. The review in the next section is how to see them before the second cycle runs.
The first three are contract failures and the fourth is a budget failure, which is the order the guide treats them in. A contract that names inputs, output, cadence and exceptions removes the first two before the dot runs once; a rule tested with an innocuous example removes the third before anything external is sent; and an owner on the allowance, with a plan for when it runs low, turns the fourth from a surprise into a line that is read every week. None of the four needs a product feature OpenAI has not announced; each needs a sentence written down before the first assignment starts.
07 — The methodThe pilot review in one page
The decision section asks for several finished assignments to be reviewed, including one where information was missing and one where an approval was required, and the review has a fixed shape. Against the contract: did every brief use only the named inputs, arrive in the stated format on the stated cadence, and stop where the contract said? Against the exceptions: when a source was unavailable, did the brief record the gap; when two sources disagreed, did it show the conflict; when a login or wider permission was needed, did the task return to the owner? Against the approvals: for each attempted consequential action, does the activity record show the attempt, the rule that caught it and the approval request, with the recipient, attachments and account unambiguous?
Against the budget: requested tasks, accepted outputs, follow-on jobs started in Codex or Work, and minutes of human correction, with the owner of the allowance named and the plan for when it is nearly exhausted written down. The budgeting section says these observations are more useful than counting messages, and the review is where they are counted.
An illustrative review for the release-page brief reads: five briefs delivered on five weekdays in the dated format; one source unavailable on Wednesday, recorded as unavailable with no substitute; one conflict between two release pages shown as a conflict; one request to send a brief to a client returned to the owner and approved by hand; two follow-on Codex tasks started, both counted; forty minutes of correction across the week, mostly on the uncertainty section. That record supports expanding one dimension, as the decision section advises, and it says which dimension: the format is stable, so the source list can grow before any action permission does.
More sources, then a longer task, then a new action permission, one at a time. A regression after a single change is easy to explain and reverse; one after three changes is a week of reading the activity record.
Start with a responsibility you can inspect
Give the dot a useful but limited job, make the approval boundary explicit and review the whole result. Persistent access becomes easier to justify when the work record shows that the agent handles both the normal path and the exceptions correctly.