OpenAI launches ChatGPT Agent

OpenAI launched ChatGPT agent, a mode selectable from within ChatGPT’s composer that combines several previously separate capabilities: Operator’s ability to act in a visual, remote browser; Deep Research’s capacity for multi-step web synthesis into written reports; and a terminal tool with limited network access for running code and analysing data. The system also connects to external applications and data sources through first-party connectors, letting it, for example, pull information from a user’s calendar or files as part of completing a task.

The company said Operator would be retired as a standalone product in the following weeks, with its functionality folded entirely into ChatGPT agent, and that Deep Research’s browser access would similarly deepen through the same unified system. The accompanying system card described the model as built on the same family as OpenAI’s o3 reasoning model, trained to plan and execute multi-step tasks with less step-by-step user supervision than earlier agentic products required, while giving users the ability to interrupt, pause or take over active sessions.

The release was part of a broader shift already visible across the industry that year: rather than treating “agent” as a separate product tier, labs were building browser and tool access directly into their flagship chat interfaces, on the premise that most users would want a single assistant capable of both conversation and autonomous action rather than switching between named tools. The system card devoted substantial space to the risks this created — prompt injection from malicious web content, and the greater consequences of a mistake made by a system that can act rather than merely advise — treating these as the main open problems for agentic products generally rather than issues specific to this release.