ChatGPT Work: Bringing the Coding Agent to Knowledge Work

The empty prompt box is a product problem

Early ChatGPT Enterprise customers arrived with deployment teams, substantial budgets and plenty of internal data. Ask what they wanted from AI, and enthusiasm came easily. Ask for a specific workplace task, and the answers diverged. Akshay recalls how a seemingly common requirement—apply AI to company information—concealed very different workflows.

[ChatGPT Work](https://openai.com/index/chatgpt-for-your-most-ambitious-work/?utm_source=chatgpt.com) brings Codex's agent capabilities to ChatGPT's audience: users can delegate tasks with access to context and tools, then supervise execution. But an unrestricted prompt box gives little guidance about what to attempt. Akshay argues that the interface must start with tasks people already have and show how AI can help. Customer-facing engineering support cannot replace that product responsibility.

For Akshay, this extends a longstanding ambition: making the capabilities of code accessible beyond engineering. An early AI testing startup and subsequent work at Airtable pursued that ambition through automation and database primitives. He came to see large language models as the missing technology. At OpenAI, he describes a culture where people can propose and ship ideas, with a persistent mission and an experimental path through products that succeed and fail.

Familiarity with chat does not automatically teach delegation. Early adopters know to supply relevant context, connect tools and supervise the agent. The next adoption challenge is making that working pattern understandable to people who already use ChatGPT but have not discovered what agents can do.

Why Codex became a productivity foundation

An unexpected internal audience helped motivate Work. In user research, Akshay encountered strategic finance and marketing staff using Codex and taking pride in their new capabilities. Agents were crossing professional boundaries earlier than the team expected. The effort described as the Merge or Super App grew from a distribution question: how could those capabilities reach ChatGPT's existing audience?

The scope also extends beyond employment. Akshay's team is called Productivity. He recounts an internal Slack story about a missing delivery: someone gave Work the courier's photograph, and the agent reportedly compared it with neighborhood listings to identify the apartment complex where the package had been left. The account establishes a reported location, not a confirmed recovery.

Codex and Work share a harness—the underlying machinery supporting the agent's tools and execution. Improvements for knowledge work, including plugins, computer use and generated outputs such as spreadsheets, apply to both. Their interfaces differ in how they expose Git state and file diffs, and in sandbox defaults that govern the execution environment. Akshay's principle is that choosing an experience should not strand a user without its capabilities.

A retirement-calculator spreadsheet makes the distinction concrete. Codex can expose file edits and diffs, with parts of its interface assuming a Git repository. Work omits those details. Both can create the spreadsheet; the product choice is how much of the underlying computer workflow the user encounters.

Shared foundations, different interaction priorities

The convergence strategy rests on a hypothesis about changing jobs. If automation lets people move between coding, strategy documents, event planning and marketing, rigid product boundaries based on professional identity become less useful. Unified plugins support movement across the experiences.

The existing Chat harness remains. Its historical priorities include low latency and personality for search, learning and conversation. Work adopts the flexible computer environment developed through Codex, while abstracting details that knowledge workers may not need.

For a spreadsheet request, Chat can encourage the user to move into Work when the model recognizes that the task is better served there. Equivalent capabilities across Chat, Codex and Work remain a future goal; their present differences reflect distinct product priorities.

Choose configuration by the task

The advice is to start with the default, then adjust if cost efficiency or output quality is inadequate. Power users can change reasoning effort and model class, but the team acknowledges that the options can be excessive. Its slider compresses several dimensions into one visible tradeoff: speed and efficiency at one end, quality and thoroughness at the other.

The task provides a useful starting point: - Ultra or multi-agent execution: complicated tasks, open exploration or work that can run along parallel tracks. - The mode transcribed as Goal: tasks where sustained progress can be verified over time. - Default configuration: most tasks, which initially fit neither special case.

Collaboration preferences matter too: some users want frequent opportunities to steer or check direction. The suggestion that a lighter model configuration with Goal could substitute for higher reasoning effort receives no general endorsement; the response emphasizes interaction effects and individual preferences.

Artifacts make results inspectable

When the conversation returns to the retirement calculator, Work has an Excel-like output visible while Codex is still running. The demonstration is not a controlled timing comparison. It illustrates the artifact viewer as an iteration workspace: users inspect a generated file as a recipient would see it, then revise it with the model.

Artifact development joined training infrastructure for better model outputs with interfaces for revising them. The team describes substantial quality improvements over versions 5.5 and 5.4, without a quantitative evaluation. The viewer's purpose is faithful rendering during revision; the response does not endorse an interviewer's suggestion that it removes the need for an Excel license.

Sharing was already available, while richer team collaboration remained under consideration. A recurring internal workflow explains the opportunity: a colleague asks a question, someone delegates it to Work, then sends back an interpretation. That handoff loses useful execution context. Yet the human intermediary may also be a permissions boundary: the colleague may not be entitled to every connected source or local file the agent consulted. Collaboration must preserve both context and access boundaries.

Sites turn answers into working surfaces

Sites lets generated websites serve as prototypes and working documents. The model-selection slider was reportedly developed almost entirely in a Site, where design, engineering and product could explore its behavior together. Corporate finance provides another internal example: recurring monthly reports moved from decks and spreadsheets into Sites as shared working artifacts.

The advantage is flexible presentation. A document tool can be constrained by its features or by what the user knows how to operate. Generated HTML allows custom layouts that foreground information otherwise buried in a deck. That freedom still requires editing: a page can contain more information than its reader can use.

One participant's Strata project follows that expansion. They report uploading roughly 30 photographs of the physical board game and obtaining a playable site with 3D block placement. Helping friends practice grew into training AI opponents through self-play—having the AIs play one another—and automated research with benchmarks. The participant reports 1.7 billion tokens, but gives no accounting method that would make this a reproducible cost estimate.

As the research accumulated too many parameters to inspect in chat, the participant requested a site exposing training settings and game-optimization work. They eventually moved it off Sites because it needed more database and infrastructure capabilities. The panel also became sprawling and number-heavy, requiring pruning. A working interface had replaced text output without automatically making the result easier to understand.

Another participant reports building a similar game in Codex in 18 minutes and 53 seconds without supplying board or rule images. That version lacked the automated research features, so the accounts do not establish a performance ranking. These personal evaluations also produce something friends can play—and show colleagues capabilities they had not realized were available.

A product that builds other products must provide enough interface to express intent and verify tools and sources, then let that interface recede. Reusable instructions, called skills, can specify presentation preferences; keeping the site beside the conversation supports further editing. Easier collaboration and shorter, better-organized outputs remain unfinished work.

Context gathering becomes useful work

The adoption sequence starts with developers, who tolerate setup and friction, then expands to general knowledge workers and everyday personal use. A saved meal plan illustrates the common mechanism: create something, retain it in the computer environment and return to it later. The Work name may discourage people from trying such personal tasks. Personalized demonstrations, based on a user's interests and previous activity, are one proposed way to help them discover more.

A practical recommendation is to retry tasks that previously failed. One speaker describes AI becoming much more useful for gathering performance-review evidence than it had been six months earlier. With access to code, reviews and Slack, it surfaced colleagues' contributions the speaker had missed.

An interviewer objects that an AI-written review can communicate a lack of care. The respondent clarifies that they would not present a review written solely by AI: their use is steerable agentic search for supporting context. They acknowledge that it probably misses information. This is a personal account of finding overlooked evidence, without a measurement of completeness or review quality.

The speaker describes a feedback loop: Codex and ChatGPT let people accomplish more, making their contributions harder for managers to track and increasing the value of AI-assisted context gathering. Saving artifacts gives later tasks a growing body of reusable information, provided the agent can access it.

One participant implements this with Markdown files: Codex writes project learnings into separate notes, and a global setup retrieves them across projects. It sometimes resurfaces an observation from four months earlier or identifies work already completed. Another internal automation reads launch discussions across Slack and documents, then posts memes built from unexpected connections. Both anecdotes show the value of finding information the user did not think to request directly.

Distribution still requires discovery

The July 2026 conversation celebrates a reported ten-million-user milestone, without an activity window or counting methodology. The interviewer says the count shifted from Codex alone to Codex plus Work; the response does not explicitly confirm that definition. At the time, Work was available only to paid users, and ChatGPT users were not placed into it by default. Reaching the broader audience still required demonstrating value, encouraging trials and learning from feedback.

The team intends to keep improving Codex specifically for software development. The broader experience should increase its utility by letting someone move between a diff and an artifact. Even artifact is internal vocabulary: users should be able to ask for the concrete output they need, such as slides.

Persistent computers and the time to gather data

A household-agent anecdote exposes an operational constraint. The respondent and their wife gave an agent a calendar and found its event creation useful. The experiment ended when the laptop running it died and was never restarted. Work on web and mobile provides a persistent computer environment: files survive between sessions, scheduled tasks can run, and stored information can be revisited. A team member reportedly moved workout planning and meal tracking into Work using these primitives.

Asked whether Work would completely replace OpenClaw, the respondent says they do not expect that, sees continuing value in its open-source technology, and cannot speak to its roadmap. Work's ambition is one conversational session extended through plugins that connect the tools a task needs, while continuing to support specialized products.

Financial planning offers a concrete case. The respondent says the finances plugin has replaced planning and budgeting functionality for their own use, including retirement planning. ChatGPT does not hold their cash or assets. With the relevant data connected, conversational questions can offer analytical flexibility without adding more filters and search controls to a conventional budgeting interface.

The interviewer questions whether fetching data from external services only when needed is sufficient, suggesting a warehouse or cache. The response turns on access patterns: immediate answers are difficult when many sources must first be queried, but delegated tasks can tolerate time spent gathering information. Programmatic tool calls and sub-agents—agents assigned parts of a larger task—can parallelize some of that work. The discussion offers this tradeoff without disclosing Work's data architecture or measured latency gains.

Delegation needs selective visibility

Sub-agents introduce an interface problem: how much activity should users see? Work limits detail by default and offers expandable transcripts. The design aims to communicate that parallel work is happening without confronting users with every agent's activity at once.

Power users may want more deliberate delegation. One participant routinely prompts for sub-agents, seeking parallel execution and possible savings from smaller models. Recurring podcast preparation provides an example: extensive research about people could assign web searches to cheaper, faster models. This steering can happen through prompts rather than additional interface controls. The discussion establishes workflow preferences, not measured savings.

The configuration has usage consequences. After launch, Ultra moved behind explicit opt-in in advanced settings. The stated purpose was to target power users who understood its behavior and that, depending on the task, it could consume more of their allowance.

Memory must learn context and surface it well

Cloud conversations in Work inherit ChatGPT memory by default and can write back through the same memory V3 system. That carries personal context from previous ChatGPT use into Work. Even short, apparently isolated chats can accumulate into a durable representation of a person.

Memory quality involves both learning useful information and bringing it up at an appropriate moment. The respondent does not work directly on memory and does not identify whether missed relevant memories or irrelevant retrievals are the larger problem. Cross-project and collaborative memory also remain areas of development, with the hope of extending existing primitives.

Chronicle is an experimental additional input to memory that learns from computer activity. It is not expected to capture everything or fully recover intent; its proposed value is surfacing overlooked information at useful moments. The distinction is between creating memories and processing existing memories: a scheduled cron job can review logs and derive skills, while Chronicle adds the underlying memories themselves.

Measure complete experiments, not generated activity

Faster implementation can shorten the path from an idea to user feedback, while broader individual capability makes more ambitious projects possible. Asked whether this means smaller teams, the response emphasizes expanded scope. The predicted shape is T-shaped capability: people work across more functions while retaining and deepening a specialty.

That shifts attention toward ideas and taste. One participant uses a front-end design skill to generate four substantially different options, then selects and combines preferred elements, acknowledging the token cost. Another reports that an automation intended to deliver new ideas has not worked for them. Useful product ideas, they argue, grow from user conversations, observed friction, feedback and existing plans. Generalists still have to connect that context to what gets built.

The productivity ambition is to let people build and interpret data beyond the limits of their role or training, and ultimately create time for personal pursuits. Measuring that remains unresolved. The proposed target is whether a user achieves a goal, but goals vary and may not be known in advance. A thumbs-down could concern content, tone or failure to make progress. Commits, lines of code, story points, token consumption and pull-request counts are activity proxies whose relationship to outcomes may weaken as generation becomes easier.

For managers, the recommended unit of practice is a complete product experiment: 1. Generate an idea and build it. 2. Collect feedback and respond to it. 3. Validate or invalidate the hypothesis. 4. Move to the next idea. The aim is to improve both the number and quality of these attempts. Efficient iteration depends on culture as well as implementation: teams need the humility to move on and the motivation to repeat the process.

The closing warning is conflating motion with progress. AI makes activity easier to produce. A team still has to define what it is trying to achieve and what evidence would show improvement. Measuring productivity across a diverse product audience is difficult; defining progress for a particular team's work remains a responsibility the tooling cannot remove.