What a dot should leave with the morning’s dependency patch

The engineer opening a dependency patch in the morning has a small, reasonable question: did the tests use the new version? Answering it should take a glance at the handoff, not another investigation. A persistent assistant is useful here if it carries the work across conversations with enough context for the next person to pick it up.

OpenAI introduced Dots on September 29, 2026 as a persistent agent experience powered by GPT-6 Astra. Each dot has a cloud computer and can work through connected apps. OpenAI’s launch announcement described an ecosystem of more than 4,000 apps and previewed specialist dots for focused enterprise pilots. That large catalog is not a promise that a particular connection can perform the operation a team needs.

For a platform team, the appealing part is less dramatic: keep the reason for a change, the affected files, attempted checks and unresolved decisions together until the next conversation. We will follow a hypothetical dependency update to see what that handoff needs. Its authorized endpoint is a reviewable patch, not a release. Leaving maintenance with an assistant does not quietly add permission to deploy it.

Give the test a workspace that can answer the question

The model can interpret the request, but files and installed software live on the working computer, and connected apps supply particular permitted operations. Continuing context helps keep the assignment available. The GPT-6 Astra documentation describes reasoning, coding, computer use, research and document creation; none of those capabilities tells us which dependency a specific test loaded. For that, we need to know the workspace.

A dot can start work in an existing Codex cloud environment, separate from its own general-purpose cloud computer. The prepared coding environment supplies the repository setup and tools for that task. Access to a local computer is optional and starts disabled. According to Getting started with your dot, setup is now available in the latest iOS or Android app, the desktop app, or a desktop browser; mobile web is not supported.

This distinction matters before the first edit. If the patch is prepared where the team’s tests can run, the eventual result is easier to assess than a proposed fix from an unspecified machine. The Codex Cloud environment explainer explains how prepared setup and individual task files stay separate. Choosing where the work runs is part of choosing what a passing check can establish.

What was the assistant actually allowed to do overnight?

“In the background” describes timing, not permission. How OpenAI describes Dots safeguards distinguishes proactive research from action-taking work. Proactive research reads permitted connected sources and produces private notes. Its tools cannot directly send messages, alter connected-app content, or control a browser or desktop. Acting on a useful finding still goes through the normal rules and checks.

An already authorized task can also continue between conversations, including preparing changes within its scope. That is different from read-only research, even if both run while nobody is chatting. The person reviewing the morning handoff needs to know which assignment was active and which operations it allowed.

A status saying “still working” leaves quite a lot to the imagination. A named patch, its test result and the decision still waiting for an engineer give the next person somewhere to begin.

The test passed. Then someone read the version number.

For our hypothetical pilot, an engineer supplies an approved nonproduction repository, the outdated dependency, a target branch and the existing test command. The assignment is to prepare the update for review. The dot returns a patch with a passing test, which looks like a useful start to the morning.

Then the reviewing colleague checks the installed dependency. It is still the old version.

The patch says what the repository should install. The running environment shows what the test used. Editing the first did not establish that the second changed. Install the version required by the patch, rerun the check, and keep its command, loaded version and tested revision together. Another green result against the old dependency would not answer the new question.

One version number has changed the meaning of the handoff. The colleague should not need to rebuild the entire workspace merely to discover that fact. A useful patch carries these details with it:

Part of the work	What travels with it	Question for the reviewer
Inspection	Dependency version, affected files and relevant issue links	Do these sources describe this repository and request?
Preparation	A patch tied to an identified base commit	Does the diff stay within the requested update?
Testing	The command, exit status and complete relevant results	Did these checks run against the proposed patch?
Review	The proposed change with unresolved failures stated	Does it meet the team’s existing acceptance criteria?

The assignment still stops at a reviewable patch. Merge and deployment belong to the engineer’s decisions. If the target branch moves during the work, reassess the patch against that new state instead of treating the earlier review as timeless. The production-agent execution guide develops this bounded-action approach. A Custom Rule can state the limit; repository permissions or protected-branch rules can enforce it independently.

Now suppose the engineer separately authorizes a pull request, but the creation request times out. Look for the PR at the destination before trying again. There may already be one waiting, even though the caller never received confirmation. The unknown-write reconciliation guide follows that uncertainty. For this trial, stay with a prepared patch whenever the target state or permitted scope cannot be established.

Continuity is doing useful work when the next person can see both the inconvenient old-version result and the check that replaced it. They inherit a review, not a mystery about what happened yesterday.

Disconnecting the ticket app does not unlearn the ticket

To explain why the dependency needs updating, imagine the engineer includes a made-up ticket. Pairing that synthetic ticket with an approved test repository makes a sensible first trial: the team can learn how context carries forward before introducing sensitive incident records.

Check the connections already available before adding another. Plugin permissions are shared across Dots, ChatGPT, ChatGPT Work and Codex, so the trial may inherit access from other work. Custom Rules can set action limits, and Auto-review checks certain proposed steps against instructions and safety requirements. The privacy, security and safety FAQ distinguishes actions that may proceed, need approval or require a person to take over.

These controls answer whether the work is permitted. They do not make a test against the old dependency validate the new one. OpenAI describes sandboxed cloud workspaces, with action-review safeguards outside the environment a dot can modify. A sandbox limits access; an allowed edit can still be wrong. The reviewer needs both answers: was the action authorized, and did it accomplish the requested change?

At the end of the trial, the engineer disconnects the ticket app. That stops new access through the connection, not information already incorporated into the dot’s context. The FAQ says individual dot memories cannot currently be viewed or edited directly. Deleting the dot’s context, managing separately stored files and managing ChatGPT memories are distinct operations. One disconnected app does not mean every retained copy has disappeared.

The synthetic ticket has served a second purpose. It lets the team learn what persists and what cleanup requires without putting a real incident record through that experiment. If organizational rules exclude the data a proposed task needs from the environment, choose a different first task.

Make the pilot fit the account that will run it

Access depends on plan, region and administrator settings, not just the announcement. The rollout covers Pro, Business Premium and Enterprise offerings. The getting-started page excludes the EEA, Switzerland and UK from the initial Pro rollout and requires administrator enablement for the Enterprise beta. It also says access can take several days to arrive.

The usage descriptions need to be read with dates attached. At publication, the announcement said conversations did not use ChatGPT limits while delegated Codex and ChatGPT Work tasks counted normally; the help page described a first-month allowance concession. Those pages still distinguish conversation access from deeper-work allowances in the November 2 check. Consult the account’s current terms before planning recurring maintenance. Neither description promises unlimited delegated tasks.

Specialist dots are a separate enterprise-pilot offering with organization-managed identities and access. A demonstration of a specialist’s capabilities does not establish that a personal dot has the same tools or permissions. Build the trial around the access the team actually has.

Leave the next engineer a shorter morning

Choose maintenance the team already knows how to review, and keep its existing test and release criteria. Count preparation, review and correction time alongside the useful work completed. Hours of background activity are not the thing the colleague came to accept.

The handoff can now say which revision was tested, which dependency was installed, what passed and what still needs a decision. If the next engineer spends less time reconstructing that story and correcting the work, persistence has saved something real. The morning begins with the patch, rather than with a search for where last night left off.