Akshay Nathan says OpenAI is extending Codex from developers to knowledge work and personal agents

This Recap covers a Latent Space podcast interview with Akshay Nathan, who leads Core Product Engineering at OpenAI. It is about OpenAI's attempt to extend the agent architecture behind Codex into ChatGPT Work for knowledge work and personal productivity. It matters because OpenAI is combining persistent computers, plugins, artifacts, Sites, memory, scheduled tasks, and sub-agents into a general-purpose work interface—and Nathan says its success should be judged by completed goals, not AI-generated activity.

OpenAI's current product guide describes ChatGPT Work as a place to delegate substantial tasks that produce reviewable outcomes. It says people who used Codex for non-coding work can stay in Codex or use the same core capabilities through an experience designed for everyday work. Nathan's interview explains why OpenAI made that choice and where it wants the product sequence to go next.

Nathan says the earlier enterprise challenge was not simply giving companies access to a capable model. Different teams wanted different outcomes, and an open-ended box made the product flexible while leaving many users unsure what to do. OpenAI needed to meet people inside concrete workflows and show them how AI could provide leverage.

The turning point came from Codex. Nathan says strategic-finance, marketing, and other nondeveloper teams inside OpenAI began using it for their own work. Their pride in using a developer product—and the feeling that it gave them a new capability—suggested that the agent model could reach ChatGPT's much larger audience earlier than OpenAI expected. ChatGPT Work grew from that observation.

Nathan draws a clear architectural line: Codex and ChatGPT Work share the same agent harness and underlying capability set. Work on plugins, computer use, and artifacts benefits both experiences. OpenAI is not building an unrelated knowledge-work agent beside its coding agent.

The products still make different interface choices. Codex can show Git state, diffs, and file changes because those details help software developers verify work. ChatGPT Work hides more of that machinery and uses different sandboxing defaults. Nathan says the products can remain opinionated without forcing users to choose a permanent occupational box, because AI is already blurring code, strategy, design, operations, and other functions.

Nathan's advice is to begin with the default. OpenAI exposes model and reasoning controls for power users, but he says most people should change them only when they encounter a concrete quality, cost, speed, or thoroughness problem. The product's reasoning slider tries to compress several technical choices into a single understandable trade-off.

He separates task shapes rather than prescribing one advanced mode for everything. Ultra or multi-agent work can help with unusually complicated, open-ended, or parallelizable tasks. Goal mode is better suited to work that can make steady, verifiable progress over time. Many ordinary tasks fit neither extreme and should start with the default configuration.

Nathan says artifacts improve the collaboration loop between model and user. Instead of accepting a block of text or an opaque file, a person can inspect a high-fidelity spreadsheet, page, or other output and ask for changes while looking at something close to the final work product. He attributes the result to both model training and product design.

He sees Sites as more than a prototyping tool. One OpenAI corporate-finance team, he says, moved recurring reports from decks and spreadsheets into interactive Sites. HTML can combine explanation, controls, data, and visual hierarchy without requiring the user to master every feature of PowerPoint or Excel. The trade-off is that richer connected context also raises collaboration and permissions questions, especially when plugins and local files expose information that not every recipient should see.

Nathan describes a three-step sequence. Developers adopted agents first because they tolerate setup and friction. General knowledge workers come next, which requires clearer interfaces plus capabilities such as artifacts and computer use. Lessons from that stage can then move into personal productivity.

The interface must show enough for users to supply context, confirm tools and sources, and verify results without displaying every possible capability at once. Nathan recommends broadening the set of tasks people attempt and retrying work that failed several months earlier. He also says more context makes the agent more useful over time, but his performance-review example preserves a human boundary: the agent gathers evidence across code, reviews, and Slack; it does not replace the manager's judgment or write the final review by itself.

The hosts describe a combined milestone of ten million Codex and ChatGPT Work users, and Nathan treats it as evidence that the product direction is working. He also calls it small beside ChatGPT's broader audience. At the time of the interview, he says, ChatGPT Work was limited to paid users and required an education, trial, and feedback process rather than an automatic switch for everyone.

Merging the capability layer does not mean ending Codex. Nathan calls it a durable brand and says OpenAI intends to keep improving the developer experience. The goal is continuity: someone can move from code to an artifact or a search across a repository without changing to a weaker agent.

Nathan says OpenClaw demonstrated the appeal of an agent that persists across work and home life. He used it with his wife to manage household calendar tasks, while a colleague later moved workout planning and meal tracking into ChatGPT Work. Persistent files, scheduled tasks, and reusable context make those use cases possible without starting from zero each time.

OpenAI's larger aim is an extensible conversational interface. Plugins can connect finances, science tools, and other specialist systems while the user asks for an outcome in one place. Nathan does not say that ChatGPT Work will replace OpenClaw or every purpose-built product. He argues instead that the core experience should absorb as much useful functionality as possible while specialist tools continue to provide best-in-class capabilities.

Sub-agents create an interface trade-off. Showing them helps users understand that a complicated task can run across several parallel tracks. Showing every step can flood the screen with detail. Nathan says OpenAI currently exposes enough activity to signal the capability while hiding more of the underlying work.

Memory creates the same tension at a longer timescale. Nathan says ChatGPT Work can inherit from and write back to ChatGPT memory, allowing years of use to carry into new work. Chronicle adds screen activity as another context source. OpenAI's current Chronicle documentation calls it an opt-in research preview and warns about sensitive screen content, unencrypted local memories, rate-limit use, and prompt injection. More context can make an agent feel personal and proactive, but it also increases the cost of retrieving the wrong thing or exposing information at the wrong moment.

Nathan expects AI to make more people generalists with a specialty. A person without deep design training can iterate on a design, while a specialist can use the same tools to go deeper. As building gets easier, however, ideas and taste become constraints. He says useful product ideas usually come from users, observed friction, feedback, or a prior plan; asking a model to “bring me new ideas” does not reliably produce grounded novelty.

Cheap execution also breaks familiar productivity proxies. Commits, pull requests, story points, and tokens can all increase without moving a team toward its goal. Nathan prefers quality at-bats: complete cycles from idea to build, user feedback, validation or invalidation, and the next iteration. The warning is the interview's limit on its own everything-app thesis. An agent can make activity abundant. People still have to define what progress is.