A release is blocked. Someone can describe the bug, but that does not clear the release. The failing test has to reproduce it, the relevant code or configuration has to change, and the suite has to run again. A reviewer will eventually need the diff and the repository state, not merely a confident account of what happened.
Codex is the practical place to start. The work has to leave evidence in the technical environment: changed files, command output, test results and a diff a reviewer can inspect.
Codex grew out of that environment-facing use case: working against a codebase, editing files, running commands and tests, and leaving a change that can be reviewed.[3][4][5] The Codex app announced in early 2026 put parallel tasks, projects, worktrees and diff review into the same command-centre context.[4] A repository migration that takes days still belongs in that context if those are the things that decide whether it is done.
The arrival of Chat, Work and Codex side by side in the ChatGPT desktop app on 9 July 2026 makes the choice easier to postpone, not easier to make.[1] The labels are familiar enough that it is tempting to sort work as conversation, office work and code. That shortcut obscures the one question that actually changes the starting point: what evidence would make the next person accept the result?
A task can change shape halfway through
The same release may have begun somewhere else. Perhaps a product manager is weighing two workable launch options. The immediate work is to expose the trade-offs: who bears the cost, which assumption is fragile, and what gives way when the calendar gets tight. Chat is a sensible place for that conversation. Its useful output can be a better judgement, a sharper question or a decision that has finally been made explicit.
Once that decision must travel to stakeholders, the job has changed. It may need a sourced recommendation, a clear narrative and material others can inspect. If the decision later requires a tested change to a technical system, only that portion needs to move into Codex. Passing work on in this order is ordinary: a task can keep its name while the evidence required to finish its next stage changes.
Work incorporates Codex technology,[1] and OpenAI has also positioned Codex for analysis, reports, dashboards, executive materials and prototypes, alongside programming work.[2] The overlap is why this article treats the finish line as a starting heuristic, not an OpenAI taxonomy of exclusive lanes.
Work begins at the handover point
Take a monthly business review. The figures may sit in a spreadsheet; the context may be spread over meeting notes and other connected applications. Leadership does not need a passing explanation of one metric. They need a deck and supporting analysis that can be reviewed, questioned and used to decide what happens next.
That makes Work a strong first stop. When it launched, OpenAI described Work as an agent that can act across apps and files to turn a goal into finished work. The announcement also described complex projects running for hours and producing sheets, slides, documents and web apps.[1]
The spreadsheet itself is not the criterion. A data-processing repair inside a versioned repository may also involve a spreadsheet, yet it asks for a different sort of evidence: changed files, tests and a record of the change. Conversely, when the only live question is why a metric fell, before anyone needs a handover-ready review, Chat may be the lighter and more useful place to start.
Chat has a smaller, useful job
Chat is valuable before a deliverable exists. It gives a team somewhere to test an interpretation, add a missing constraint or notice that the original question was poorly formed. In the launch example, producing a formal deck too early would only give an unsettled argument better typography.
The boundary moves once the desired result moves. A conversation can become an editorial brief for Work; a stakeholder recommendation can expose a technical change that Codex must make verifiable. Starting in Chat does not downgrade a task. It acknowledges that, at that moment, human judgement is the thing being accepted.
Keep the execution host in view
Codex on mobile can help someone inspect or steer work running on a laptop, devbox or managed remote environment. The files, credentials, permissions and actual execution remain with that host.[6] Mobile is useful for continuity and supervision; it does not create a separate category of work.
Before opening anything, it is worth writing a short done when line. Name the thing that will remain, who or what accepts it, and where the evidence has to appear. If one of those answers changes, the surface may need to change as well. That small pause usually saves more time than choosing the button that appears to promise the most.
References
- OpenAI, “ChatGPT is now a partner for your most ambitious work” (9 July 2026)
- OpenAI, “Codex for every role, tool, and workflow” (2 June 2026)
- OpenAI, “Introducing Codex” (16 May 2025)
- OpenAI, “Introducing the Codex app” (2 February 2026)
- OpenAI, “Unrolling the Codex agent loop” (23 January 2026)
- OpenAI, “Work with Codex from anywhere” (14 May 2026)