OpenAI’s ChatGPT Work Launch and the Shared Agent Harness with Codex

TL;DR

OpenAI launched ChatGPT Work to bring the agent power of Codex to knowledge workers, merging the two experiences through a shared harness so users do not need to choose between interfaces. The product provides persistent computers, artifacts, Sites, memory, and sub‑agents, while the team argues that as building becomes easier, ideas and taste become the bottleneck and progress must be measured by quality at‑bats rather than raw motion.

Why Codex Spread Among Non‑Developers at OpenAI

Akshay observed that after Codex was released internally, non‑developer teams such as strategic finance and marketing began using it and felt proud to do so. He quoted a common reaction: “It's like I'm not supposed to be using it but I am.” Users described the experience as gaining a new superpower, which signaled that the agent’s usefulness extended beyond software engineers.

Product Insight Behind ChatGPT Work

The insight came from seeing Codex adoption across roles and realizing that the power of agents should be available to the existing ChatGPT user base without forcing them to switch tabs or learn a separate product. The goal was to bring the same capabilities into the ChatGPT interface so that anyone doing work‑related tasks could access them in the place they already used.

Shared Agent Harness and UX Differences

Akshay confirmed that the underlying harness— the set of capabilities that lets the agent use plugins, computer use, artifacts, and sandboxing—is identical in Codex and ChatGPT Work. Improvements made for knowledge work (better plugin support, computer use, artifact quality) are available in both modes. The differences are opinionated UX choices: in Codex mode the UI shows Git diffs and repo state by default, while in Work mode those details are hidden to reduce clutter. Sandboxing defaults also differ, but the core capabilities remain the same.

Why OpenAI Merged the Experiences

The team merged Codex and ChatGPT Work because AI is blurring the boundaries between engineering, design, strategy, and operations. Akshay argued that drawing a hard boundary based on job title would box users in, whereas a unified experience lets people move fluidly between writing code, creating artifacts, and collaborating on Sites without needing to decide which "mode" they are in. The shared primitives (plugins, computer, file system) enable this fluidity.

Model Configuration Advice for Power Users

OpenAI chose a default model configuration that it believes works best for most users. Akshay advised that most people should stick with the default and only change it if they notice a lack of efficiency or quality. For power users, options exist to increase reasoning (Ultra), use multi‑agent setups, or adjust the goal‑based slider that trades speed for thoroughness. He noted that the default should be good enough for the majority of use cases.

Artifacts, Agentic Spreadsheets, and Sites

A major push for the launch was improving artifact quality— making spreadsheets, documents, and other outputs high‑fidelity enough to share with coworkers. Akshay showed a retirement‑calculator spreadsheet that looked like Excel without requiring a license. He also highlighted Sites as a flexible, hosted webpage that can replace decks, spreadsheets, and traditional reports because it can render any HTML and support real‑time collaboration. Teams are already using Sites for month‑to‑month reports instead of slide decks.

Persistent Computer Environment and Personal Agents

Inspired by internal experiments with OpenClaw, ChatGPT Work provides a persistent file system and scheduled tasks so that agents can continue work across sessions. Akshay described how his wife used OpenClaw to manage household calendars and how similar workflows now appear in ChatGPT Work for meal planning, workout tracking, and financial budgeting. The agent can reference stored files over time, enabling personal‑productivity use cases that blend work and life.

Sub‑Agents and Ultra Mode

The product shows when sub‑agents are used, but the UI hides the detail by default to avoid overwhelming users. Akshay explained the trade‑off: displaying sub‑agent work gives transparency but can be noisy, while hiding it keeps the interface simple. Power users can expand the view to see what sub‑agents are doing, steer them to cheaper models, or inspect cost and latency trade‑offs. Ultra mode and multi‑agent setups are reserved for tasks that are highly parallelizable or require deep exploration.

Memory and Personalization

Memory and Personalization ChatGPT Work inherits the V3 memory system from ChatGPT, allowing the agent to recall past interactions and write back to memory. Akshay noted that memory feels personal and helps the agent surface relevant context proactively. He also mentioned Chronicle, an experimental feature that logs computer usage to enrich memory with insights from Slack, documents, and local files, though it is not enabled by default.

Impact on Product Development and Team Roles

Akshay reflected that AI has collapsed the traditional boundaries between product management, engineering, design, and operations, making it easier for individuals to act as generalists with a specialty. He argued that as almost anyone can build, the bottleneck shifts to ideas and taste, and that LLMs still struggle to follow the instruction “bring me new ideas.” Consequently, measuring productivity must move away from proxies like commits, tokens, or pull requests toward quality at‑bats— the ability to go from idea to build to feedback to validation repeatedly. He warned against conflating motion (increased activity) with real progress, urging teams to be deliberate about what progress looks like for them.

Closing Thoughts

With ten million combined users of ChatGPT Work and Codex, Akshay sees the launch as a culmination of the long‑standing vision to bring useful agents to everyone. He expects the next phase to involve scaling to a hundred million users and continuing to refine the balance between simplicity and capability, ensuring that the default experience remains powerful while preserving advanced options for those who need them.