OpenAI Maps a Third Era of Persistent AI Coworkers, With Access and Reliability Still in the Way

OpenAI’s next workplace interface, in product lead Tara Seshan’s telling, is not a better chatbot or a one-person agent. It is a persistent AI coworker that carries shared context across longer work loops—a shift that makes access, reliability, and coordination central product problems.

Seshan frames OpenAI’s consumer-product evolution in three stages: chat first, agents working alongside users second, and persistent AI coworkers third. The final category is defined less by a single completed task than by ongoing collaboration: agents retain shared context, work over longer loops, and are jointly steered with people.

That framing changes the unit of work. Instead of one person directing one agent, Seshan described a multiplayer interface in which groups of people steer groups of agents with shared data access. The stated destination is collaborative software, rather than another isolated assistant running beside an employee.

For now, Seshan said OpenAI separates Chat for search and conversation, Work for knowledge work, and Codex for development. Work uses Codex underneath while removing coding-specific surfaces, including worktrees and chain-of-thought detail. That arrangement makes a technical agent more approachable for non-developer work, while preserving distinct interfaces for different jobs.

Seshan’s stated north star is a single prompt box that chooses the right model and operating harness—the layer that determines how an agent uses tools and carries out work—without asking users to toggle modes. OpenAI shipped Work inside ChatGPT on web and desktop before waiting for full polish, according to the account of her remarks. The approach favors utility and iteration while the underlying interfaces are still being simplified.

Seshan said product teams should build for model capabilities expected two to three months ahead. Building only for the models available today would leave products behind the capability curve, she argued, while designing a year ahead would be too speculative. The result is a compressed planning cycle in which product teams must stay closely aligned with research and be ready to revise their assumptions quickly.

What the coworker model requires

• Longer-running agents with shared context and collaborative steering, rather than a sequence of disconnected prompts.
• Shared data access for multiple people and multiple agents working on the same task or workspace.
• Reliable access to local data, cloud infrastructure, and the workplace systems where the relevant work actually lives.

Seshan identified local data access, cloud infrastructure, reliability, and system access as practical blockers for persistent agents. Her comparison was a new hire locked in a room: an agent without access to documents, communications, or databases cannot contribute much, even if its reasoning improves. The product challenge is therefore to connect an agent to the work environment without losing dependable operation as that environment becomes more complex.

The vision also leaves a consequential design question inside the word “multiplayer.” Shared data access and group steering describe how coworkers might collaborate with agents, but they also make the handling of context, access, and reliability part of the interface itself. OpenAI’s current division between Chat, Work, and Codex shows the company is still operating through separate modes as it pursues that more unified destination.