Codex changed quickly enough in 2026 that two articles written a few months apart can look as if they describe different products.
In February, the Codex app was still centred on software development, but it already let one person supervise several agent threads. By April, Computer Use, an early in-app browser, chats and memory had widened the environment in which Codex could work. In June, OpenAI published usage data showing non-developers using Codex for research, data analysis, spreadsheets, presentations, contracts and workflow automation. In July, Chat, Work and Codex were brought closer together in the desktop product while GPT-5.6 introduced a new model family for the same broader work surface.[1][5][6][7]
That makes a chronological reading more useful than a current feature checklist. This article keeps each capability inside its historical cutoff. A feature that existed in June will not be quietly inserted into an April milestone simply because it is available today.
This article uses Agent Work OS as an editorial label for that trajectory. It is not an official OpenAI product name, and Codex is not literally replacing a conventional operating system. The narrower claim is that the product increasingly manages more than model output: agents, workspaces, tools, long-running objectives, permissions, reusable workflows and work that can continue across devices.
Research verified through: 18 August 2026. This is a living article. The announced retirement of GPT-5.4 and GPT-5.4 mini on 31 August is still a future event at this cutoff, so it is described as scheduled rather than completed.
February | Agent command centre
2 February 2026 | The Codex app puts several agents under one control surface
OpenAI launched the Codex app for macOS on 2 February and described it as a command centre for agents. The app could run several agent threads in parallel, use worktrees to isolate concurrent work, and bring Skills and Automations into the same desktop workspace.[1]
The practical change is easy to see without inventing a new category. A common unit of work had been one agent handling one coding task. The app raised the control surface: one person could now watch several workstreams, while agents worked in isolated environments and returned changes for review.
This was not yet the later ChatGPT Work product. The February launch was still strongly framed around the software lifecycle. What had changed was the product problem itself. Supervising agents was now part of the interface, rather than something a user had to approximate with several terminals and chat windows.
Work unit: one coding task → several agent workstreams under supervision.
5 February 2026 | Mid-turn steering makes a running agent adjustable
Three days later, GPT-5.3-Codex arrived. Alongside stronger long-running task, research and tool-use capabilities, the Codex app added mid-turn steering: a user could send another message while Codex was working and redirect the current turn without discarding its context.[2][3]
That adds a useful control point to the old prompt → wait → result loop. Long tasks no longer had to be perfectly specified before they started. When new evidence appeared, the user could intervene without starting over.
Two smaller February changes belong beside it. On 10 February the app added clearer guidance for entering Plan mode and support for parallel approvals. On 12 February GPT-5.3-Codex-Spark entered research preview as a low-latency coding path, initially text-only with a 128k context window and separate limits.[2] Taken together, the month was already separating planning, approval and latency into different controls rather than treating every task as the same model invocation.
March–April | The agent gets a wider working environment
5 March 2026 | GPT-5.4 brings computer use and tool search into the main line
For this timeline, the interesting part of GPT-5.4 is not its benchmark table. OpenAI described it as the first general-purpose model it had released with native computer-use capabilities in Codex and the API. Codex also received experimental support for a 1M context window, alongside stronger tool use and tool search.[4]
That widened the environment the model could reason over and act within. Work was no longer limited to reading and changing a repository; an agent could plan across a longer context and a larger tool ecosystem. Later browser, computer-use and plugin workflows depended on this kind of model capability becoming normal rather than exceptional.
On 12 March, revamped Automations let users choose local or worktree execution and specify model and reasoning level. On 17 March, GPT-5.4 mini was explicitly positioned for lighter coding tasks and subagents, consuming about 30% as much of the included limits as GPT-5.4.[2] That was an early, concrete version of a pattern that became more important later: coordination and bounded support work do not always need the same model profile.
25 March 2026 | Plugins turn configuration into an installable bundle
On 25 March, Codex added plugins: installable bundles that could package Skills, app integrations and MCP server configuration, available across the app, CLI and IDE extensions.[2]
The useful part is portability. Before that, making an agent good at a particular job could easily leave knowledge scattered across prompts, personal settings, MCP configuration and local files. A plugin created a more coherent packaging boundary, so a capability could be installed, shared and reused across Codex surfaces.
16 April 2026 | The workspace starts to extend beyond the codebase
16 April is one of the easiest dates to flatten incorrectly because several capabilities landed together and then evolved again over the following weeks.
OpenAI described Codex as becoming a broader workspace for getting work done with AI. The release included Computer Use, an early in-app browser, projectless Chats, thread automations and a memory preview. The software-development side also gained a fuller pull-request workflow, multiple terminals and an alpha path for SSH remote connections.[2][5]
The browser history needs a precise boundary. On 16 April, the early in-app browser could open local or public pages, accept comments on the rendered page and let a user ask Codex to address that feedback. The explicit ability for Codex itself to operate the in-app browser, click through a rendered interface, reproduce a visual bug and verify a local fix appears in the 23 April changelog.[2]
Computer Use was already more direct: Codex could see, click and type in macOS applications. The operational environment was therefore moving beyond repository and terminal state into graphical applications and rendered interfaces.
Work unit: codebase-local task → work that can cross terminal, browser, GUI and general workspace context.
23 April 2026 | Browser Use shortens the frontend verification loop
On 23 April, Browser Use in the Codex app allowed Codex to operate the in-app browser against local development servers and file-backed pages. It could click through the rendered UI, reproduce visual problems and check a local fix. The same release added automatic approval reviews, which could route eligible approval prompts through a reviewer agent before execution.[2]
Those capabilities solve different problems. Browser Use provides another external feedback surface, reducing the hand-off between “the code changed” and “the interface now behaves as expected”. Auto-review adds a governance step around actions that cross an approval boundary. Delegating a review does not increase the agent’s authority beyond the configured sandbox and approval policy.
The browser then gained Developer mode on 11 June, providing controlled Chrome DevTools Protocol access for network traffic, console output, runtime errors and page state.[2] The cleaner historical sequence is therefore: browser and annotation surface on 16 April, agent-operated Browser Use on 23 April, and CDP debugging on 11 June.
May–June | The unit of work becomes longer
21 May 2026 | Goal mode turns a task into an objective that can keep running
Goal mode left experimental status on 21 May and became available in the Codex app, IDE extension and CLI. OpenAI’s release notes explicitly said a goal could keep Codex driving towards a specific objective for hours or even days.[2]
A goal does not remove the need for checkpoints, verification or human supervision. It changes the time scale the product is designed to carry. Once a task can survive across many turns and interruptions, state, resume behaviour, notification, approval and failure handling become part of the normal workflow rather than edge cases.
The same release added Appshots and a controlled form of remote computer use after a Mac locks.[2] Those features fit the same operational problem. When work can run for much longer, the user will often be in another application or away from the machine when Codex needs context or supervision.
Remote had already started moving in that direction on 14 May: ChatGPT mobile could connect to a Mac running the Codex app and use the projects, files, credentials, plugins, skills and configuration on that host.[2] On 26 May, GPT-5.3-Codex also stopped being a user-selectable model for ChatGPT-authenticated Codex users, while API-key workflows followed a different lifecycle.[2] That is a useful warning for any historical article that hard-codes a model name into its advice.
2 June 2026 | A coding-only description no longer matches the usage data
On 2 June, OpenAI published a report on knowledge-work use. Codex had more than five million weekly active users, more than six times the level around the February desktop launch. Knowledge workers represented about 20% of users and were growing more than three times as fast as developers.[6]
Their work included reports, spreadsheets, presentations and contracts, as well as research, data analysis, workflow automation and lightweight tools that might previously have required engineering help.[6]
This is a different kind of evidence from the April feature release. In April, the product surface suggested that the range of possible work was widening. By June, OpenAI had published usage data showing that people were actually carrying Codex into work outside software development. That is one reason the Agent Work OS framing is useful here: it describes a widening work surface and control plane, not a single model rebrand.
Work unit: software deliverable → a broader class of delegated knowledge-work outputs.
18 June 2026 | Record & Replay turns one human demonstration into a reusable Skill
Record & Replay launched on macOS on 18 June. A user could demonstrate a workflow and have Codex turn it into a reusable Skill. Initial availability excluded the EEA, the UK and Switzerland, and Computer Use had to be enabled.[2]
That is materially different from writing a longer prompt. The demonstration contains an operating sequence inside an environment, while the output lands in the reusable capability layer. For repetitive work, a human SOP gained a shorter path to becoming something the agent could execute again.
25 June 2026 | Remote reaches general availability
Codex Remote reached general availability on 25 June. From the ChatGPT mobile app, a user could start or continue work on a connected Mac or Windows host, review progress and approve actions from a phone.[2]
The important detail is that the execution environment remained on the host. The phone was another control surface, not a smaller duplicate of the entire development machine. Projects, local files, credentials and tools could stay where the work already lived while the user moved elsewhere.
The same month also opened a migration path. On 9 June, Codex added flows for importing supported setup from Claude Code and Claude Cowork. On 21 July, /import expanded to Cursor and Claude Code settings, MCP servers, plugins, sessions, commands and project-scoped memories.[2] That line continued into August.
July–August | Agents, models and capabilities converge on one work surface
9 July 2026 | Codex joins the ChatGPT desktop app as Work and GPT-5.6 launch
On 9 July, Codex became part of the ChatGPT desktop app. Existing Codex app users could retain their projects, settings and workflows. OpenAI launched ChatGPT Work and the GPT-5.6 family on the same day.[2][7][8]
The combination says more than any single feature. ChatGPT Work was presented as an agent that can act across apps and files, break a goal into steps and stay with a project for hours. OpenAI also stated that Codex technology was built into that experience.[7] GPT-5.6 introduced Sol, Terra and Luna as capability tiers, with model and effort choices available in Codex; its ultra setting was described as coordinating multiple agents across parallel workstreams.[8]
One boundary matters here. The existence of Sol, Terra and Luna does not prove that Codex automatically routes every category of task to a fixed model. The supported claim is narrower: the product now exposes clearer intelligence, latency, cost and reasoning tiers, which can be allocated differently across workloads and agents.
By this point, February’s parallel agents, Skills and Automations were sitting beside computer use, goals, remote work and plugins inside a much broader desktop environment.
Work unit: work in a dedicated Codex app → work coordinated across Chat, Work, Codex and several model tiers.
21 July 2026 | Multi-Agent V2 makes the agent topology configurable
Codex CLI 0.145.0 stabilised the opt-in Multi-Agent V2 experience with configurable subagent models, reasoning levels and concurrency, restored roles and improved agent navigation.[2]
The February app had already supported parallel agent threads. This update addressed a different layer: orchestration inside multi-agent work. Once subagents can use different models and reasoning settings, a system can more deliberately separate coordination from bounded worker tasks rather than treating “more agents” as several identical windows.
The same release expanded /import across Cursor and Claude Code, including settings, MCP servers, plugins, sessions, commands and project-scoped memories.[2] On 23 July, the desktop app added Voice coordination across Chat, Work and Codex alongside multi-folder local projects. On 30 July, Review could inspect diffs across repositories inside the same multi-folder project.[2] The workspace itself was becoming broader than one repository.
7 August 2026 | Agent Plugins start to look like a portable capability distribution layer
As of this article’s cutoff, the last main milestone is 7 August. Codex CLI 0.147.0 added portable Agent Plugins and discovery across local, personal, workspace and remote plugin catalogues. The same release added --approve-for-me, import of Cursor-managed skills and synchronisation of imported Claude and Cursor conversations.[2]
This extends the March plugin idea. March established a bundle for Skills, integrations and MCP configuration. By August, the system was also dealing with where those capabilities are discovered, installed and carried between scopes. For a long-lived agent environment, that matters because capabilities that only exist as one person’s local configuration become a new form of configuration debt.
On 31 July, OpenAI had also announced that GPT-5.4 and GPT-5.4 mini would leave ChatGPT-authenticated Codex on 31 August, with GPT-5.6 Terra and Luna recommended as replacements. API and API-key-authenticated Codex follow a different lifecycle.[2] On 18 August, that retirement is still scheduled rather than completed.
What does “Agent Work OS” mean in this article?
Viewed as a feature list, the past six months are noisy. In chronological order, the changes are easier to separate.
The model layer gained clearer capability and cost tiers. The agent layer moved from one thread to parallel work and configurable subagents. The workspace expanded from repositories into terminals, browsers, desktop applications, remote hosts and multi-folder projects. Skills, Plugins, MCP and Record & Replay added reusable capabilities. Steering, goals, automations, approvals, review and remote continuation added more control over work that persists beyond a single response.
This article calls that combination an Agent Work OS for the purposes of this series. It is not a replacement for macOS or Windows, and OpenAI has not announced it as a formal product category. It is simply a useful way to read Codex in 2026: pay attention to how work is delegated, where it can continue, who can redirect it, which capabilities can be reused, and how results come back for review.
That also explains why this series needs historical cutoffs. Advice that was correct in February does not need to look identical to advice written in August. An April browser article should not quietly depend on Developer mode that appeared in June. Later articles can go deep on individual features and operating methods; A00 keeps the dates straight.
The rule for future updates is deliberately strict: a new event becomes a main milestone only when it changes the working model of Codex. Routine UI polish, patch releases and small additions belong in the changelog, not in a second changelog disguised as an article.
References
- OpenAI, Introducing the Codex app, 2 February 2026.
- OpenAI, ChatGPT & Codex changelog, verified 18 August 2026.
- OpenAI, Introducing GPT-5.3-Codex, 5 February 2026.
- OpenAI, Introducing GPT-5.4, 5 March 2026.
- OpenAI, Codex for (almost) everything, 16 April 2026.
- OpenAI, Codex is becoming a productivity tool for everyone, 2 June 2026.
- OpenAI, ChatGPT is now a partner for your most ambitious work, 9 July 2026.
- OpenAI, GPT-5.6: Frontier intelligence that scales with your ambition, 9 July 2026.