OpenAI engineers demo ChatGPT Agent handling multi-step tasks like scheduling, planning, and presentations
Jul 17, 2025 with Yash Kumar & Isa Fulford
Key Points
- OpenAI merges Deep Research and Operator into ChatGPT Agent, a single system that combines information synthesis with autonomous action across virtual browsers, terminals, and API connectors like Gmail and GitHub.
- Tasks run from 5 minutes to over an hour, with OpenAI engineers treating longer runtimes as a feature that will eventually scale to multi-day workflows as reliability improves.
- Pro plan subscribers gain access July 17, followed by Plus users within days and enterprise access in coming weeks, though the company acknowledges reliability at scale remains the primary engineering focus.
Summary
OpenAI launched ChatGPT Agent on July 17, merging its two earlier standalone products, Deep Research and Operator, into a single system capable of both research and autonomous action. The product gives the agent access to a virtual computer equipped with a text browser, visual browser, and terminal, trained end-to-end using reinforcement learning consistent with OpenAI's prior reasoning models.
The integration architecture is the core differentiator. By combining Deep Research's information synthesis with Operator's action-taking capability, and layering in API connectors for services like Gmail, Google Drive, and Linear, the agent can move from research to execution within a single workflow. GitHub integration is already in use internally, allowing engineers to query unfamiliar codebases in multi-turn conversations, a capability the earlier Deep Research model did not support natively.
Task duration ranges from roughly 5 minutes to over an hour, with the OpenAI engineers framing longer runtimes as a feature rather than a limitation. The expectation is that as reliability improves, the agent will be capable of handling multi-day tasks that currently require sustained human effort. Push notifications alert users when asynchronous tasks complete, addressing the unpredictability of variable run times.
On the roadmap, voice is flagged as a natural next form factor, with one engineer noting that text-based commands to the agent can already handle insurance comparison and account management tasks through the virtual browser. Password manager integration and richer Python-based data visualizations inside reports are cited as near-term areas of interest.
Rollout is tiered. Pro plan subscribers gain access by end of day July 17. Plus users follow within days, and enterprise access is expected over the coming weeks. OpenAI's engineers are candid that reliability at scale remains the primary engineering focus and that the product is still early-stage despite its broad toolset.