Akshay Nathan

Akshay Nathan is an OpenAI engineering leader involved in building ChatGPT Work, an agent experience for knowledge work. In his July 2026 interview with Latent Space, he was introduced as leading Core Product Engineering and described his own remit as the Productivity pillar. His account explains a consequential product choice: bringing Codex’s agent capabilities to a wider audience while adapting how much technical detail users see.[1]

Nathan attended internal user-research sessions where strategic finance and marketing staff were already using Codex. He recalled being struck by their pride in using it and their feeling that they had acquired a superpower. That experience helped motivate the team to bring agents to people who already used ChatGPT: Codex’s usefulness beyond software development had become apparent earlier than they expected.[1]

His earlier work on ChatGPT Enterprise informed the adoption problem. Nathan said enterprise customers’ concrete use cases varied enormously, even when their initial enthusiasm sounded similar. He argued that helping users discover what to do with AI is a product responsibility as well as a deployment task: the interface should meet people at the problem they are trying to solve.[1]

Nathan described Codex and ChatGPT Work as sharing the same agent harness and underlying capabilities, including improvements for plugins, computer use and artifacts. The team differentiated their interfaces through Git visibility, file diffs and sandbox defaults. His example was a retirement-calculator spreadsheet: in Codex, with the appropriate repository context, users could see file edits and spreadsheet diffs; Work hid that presentation. The flexible computer environment remained useful for knowledge work even when its technical details were abstracted away.[1]

The team considered keeping the experiences separate. Nathan explained their decision to bring them together through a shared hypothesis that AI would blur boundaries between coding, strategy, marketing and other functions. Common primitives would let users move between experiences without assigning them permanently to an occupational category. As he put it, “we should enable users to choose, but we shouldn’t box them in.”[1]

Nathan advocated a strong default model configuration, recommending that most users start there and adjust it when cost, efficiency or quality requires a different choice. He described the model slider as reducing several configuration dimensions to one axis, from speed and efficiency to quality and thoroughness. Design, engineering and product collaborated on the slider almost entirely through a Site, using an interactive prototype to explore how the control felt.[1]

He described artifact improvements as joint work with research teams: training infrastructure helped models produce better files, while product viewers let users inspect and iteratively refine them. Nathan also saw Sites as work products rather than only prototypes. He recalled corporate finance colleagues using Sites for recurring reports previously assembled in decks and spreadsheets, with the website becoming their collaboration medium.[1]

The same balance between simplicity and capability shaped subagent visibility. Nathan said the team wanted to show that complicated tasks could be handled along parallel tracks, while avoiding an overwhelming display of every agent’s activity. By the July interview, Ultra had been moved behind an explicit opt-in in advanced settings for power users, partly because it could consume more of their usage limits.[1]

Nathan’s interest in personal agents included a household experiment: he and his wife set up OpenClaw with a calendar, and it began creating events for them. They stopped using it after the laptop running it died. He described OpenClaw as an inspiration for ChatGPT Work’s personal-productivity uses, supported by persistent files and scheduled tasks. He distinguished that inspiration from knowledge of OpenClaw’s roadmap, which he said he could not speak to.[1]

Nathan described using AI to gather context for employee reviews, including contributions visible in code, reviews and Slack that he might otherwise miss. He drew a clear boundary around that practice: he would not present a review written solely by AI. Its value, in his account, was gathering and steering through information rather than replacing his judgment about a person.[1]

He predicted that AI would enable more people to become generalists while retaining a deep specialty, and that ideas and taste would increasingly constrain what teams could build. Yet he said an automation that simply brought him new ideas did not work as he wanted. Useful product ideas, he argued, arise from user conversations, observed friction and feedback; builders still need to close that loop.[1]

Nathan argued that productivity should be measured by whether people achieve their goals, while acknowledging that measuring this across diverse users remained unsolved. He viewed tokens, pull requests and other activity proxies as increasingly unreliable with AI. For teams, he emphasized quality “at-bats”: generating an idea, building it, collecting feedback, revising it and validating or invalidating the hypothesis. His warning was “conflating motion and progress”—more output only helps when a team has a deliberate view of what it is trying to achieve.[1]

Before his OpenAI work, Nathan worked in consumer fintech and automated testing. At Wealthfront, he was a technical lead on a project to build a brokerage and banking stack from scratch. He then cofounded walrus.ai, an automated testing company. Asymmetric Capital’s biography says Airtable acquired walrus.ai in 2021; Nathan subsequently became an engineering leader focused on Enterprise at Airtable.[3][1]

Nathan recalled trying to automate testing with AI before LLMs and vision LLMs, then working at Airtable to make database primitives accessible to more people. Looking back in July 2026, he described LLMs as the missing technology for bringing the capabilities of code to people without requiring them to understand its implementation. That was his own explanation of how the earlier work supplied context for ChatGPT Work.[1]