Don't forget about the UI for AI

Andrew Ambrosino says OpenAI wants one place where users can type or speak a request and have the most capable models and agents handle anything from software development to knowledge work and quick answers. ChatGPT has more than a billion consumers, while Codex’s coding-agent app grew rapidly earlier in the year. ChatGPT Work brought Codex’s harness to a broader audience, but the products remain separate tabs because a quick search should not trigger a long Codex reasoning process. OpenAI will merge the experiences only when it can provide a continuous spectrum without forcing users to choose a mode.

Ambrosino argues that local computer access makes the desktop app useful even though most software now runs in browsers. Codex can work with files, a Git checkout, local Node.js calls, and a browser for testing—capabilities constrained by browser security and sandboxing. OpenAI is also bringing cloud Codex into the browser, while the desktop app can act as a bridge to the user’s own computer and browser when an existing checkout is available. He says sandboxing remains important and that OpenAI has additional work in progress around desktop security.

OpenAI’s new ChatGPT Space is intended as a collaborative workspace where people, their agents, and shared artifacts can work together in real time. Rather than putting everyone into one chat—with competing memories, connectors, and permissions—each person keeps an agent with personal tools and preferences, then collaborates on pages, sites, and documents. Ambrosino says this avoids reducing the group to the lowest common denominator of available connections. Teams can plan a feature on live pages, tag an agent to research weekly users, insert an interactive chart, and revise it collaboratively.

To support character-level document editing, live cursors, brief disconnections, and offline states, the team built a full CRDT system in five or six weeks. The harder product questions involved agents editing alongside people: whether multiple tagged requests should run in parallel, how their threads should be organized, and how everyone sees updates. Ambrosino compares the solution to Codex’s fluid computer-use cursor. The cursor is not necessary for the model, but its motion helps users understand what is happening; similarly, animated document changes let people feel which text was edited instead of watching content jump unpredictably.

Ambrosino says OpenAI launched more than 20 things at Dev Day while simultaneously redesigning Spaces, the desktop sidebar, and the browser version of ChatGPT. He believes the speed of development creates a new problem: users and organizations cannot absorb a new feature every day, even if agents can build features quickly. Judgment and taste determine which explorations belong together and which are not ready. OpenAI tries to test products internally first; if employees do not love something, the team does not rationalize that another audience will. He also says the terminal remains important, including in a new Codex CLI refresh.

The Codex product, research, and core-harness teams work closely on features such as code mode, where the model writes code to plan several computer-use steps instead of reasoning through every action individually. Plugins now provide a cleaner extension mechanism than stuffing behavior into developer prompts. For safety, an approval model judges whether a tool action is safe and matches the user’s request; settings can approve or block applications and websites, while some delegation capabilities remain restricted. Spaces itself was developed inside its own pages using interactive visualization blocks, letting the team generate mock-up variations, test single-player or multiplayer behavior, and tag Codex to build the chosen design.