ChatGPT’s new Agent capability is genuinely different from what I expected.
The agent’s operations are rendered in real-time on screen with full verbosity — you’re watching it work, and you can trace exactly why it did each thing. It’s less like using a tool, more like watching a capable human assistant operate a workstation on your behalf.
The team building it came from ChatGPT’s Deep Research and Operator teams — and one of the researchers described it as “a better version of Deep Research + Operator fused together.” That framing tracks with what I’m seeing.
For anyone thinking about building something similar with open-source: the key architectural insight seems to be that the agent’s reasoning needs a fast deep-research-style retrieval foundation. The difference between a good agent and a great one might live in that layer.
💬 Read more: 2025 Year in Review (English)