🤖

ChatGPT Agent Launch Summary

Jul 17, 2025

Summary

  • The meeting introduced and demoed the new ChatGPT Agent, a unified AI agent capable of complex, multi-step real-world tasks using a virtual computer equipped with text and visual browsers, a terminal, and integration with external services.
  • The team reviewed the agent's features, collaborative workflow, tool selection, and security considerations, with multiple live demonstrations including event planning, swag ordering, and automated evaluations.
  • The ChatGPT Agent launches today for Pro, Plus, and Team users, with wider rollout for Enterprise and EDU users expected by the end of the month.
  • Attendees included Yash, Jing, Casey, Issa, and Sam, representing product, research, and engineering teams.

Action Items

  • By end of day – Product Team: Complete rollout of ChatGPT Agent for Pro, Plus, and Team users.
  • By end of month – Product Team: Launch ChatGPT Agent for Enterprise and EDU users.
  • ASAP – Communications Team: Update documentation and user guidance to include new security warnings and agent usage best practices.
  • Ongoing – Engineering Team: Continue monitoring for prompt injection and other novel attack vectors; update safeguards as needed.
  • Ongoing – Research Team: Gather early user feedback to inform rapid iteration and improvements.

ChatGPT Agent Launch and Demo

  • Launched ChatGPT Agent, a unified AI agent combining capabilities of previous "Deep Research" and "Operator" tools.
  • The agent can autonomously use a virtual computer with text/visual browsers, a terminal, access to APIs, and generate/modify documents, slides, and spreadsheets.
  • Demonstrated real-world scenarios including wedding planning, gift and outfit recommendations, hotel booking, swag ordering, and itinerary creation.
  • Live demos highlighted agent's ability to:
    • Take multi-step instructions
    • Clarify and ask for user input
    • Be interrupted mid-task and receive additional instructions
    • Collaborate and confirm before taking sensitive actions (e.g., purchases)

Technical Approach and Model Training

  • Reinforcement learning used to train model on hard tasks requiring multiple tools and decision-making about tool selection.
  • Agent trained for collaborative, interruptible, multi-turn workflows, allowing user intervention and mid-task clarifications.
  • Benchmarked against internal and industry-standard tasks, showing strong performance in intelligence, agentic, browsing, spreadsheet, and banking benchmarks.

Security, Trust, and Rollout Considerations

  • Highlighted potential security risks, especially prompt injection attacks, and described layers of defensive monitoring and real-time updates.
  • Emphasized importance of user caution with sensitive data, using “takeover mode” for confidential information.
  • Rollout plan:
    • Available immediately for Pro, Plus, and Team users (Pro: 400 queries/mo, Team: 40 queries/mo).
    • Enterprise and EDU access expected by month-end.

Decisions

  • Proceed with ChatGPT Agent launch for Pro, Plus, and Team users today — based on product readiness and successful integration of previous tool functionalities.
  • Rollout to Enterprise and EDU users by end of month — dependent on further monitoring and updates.

Open Questions / Follow-Ups

  • Monitor effectiveness of prompt injection defenses and gather real-world attack data post-launch.
  • Assess user understanding of new agent workflow and guidance—determine if further documentation or education is needed.
  • Gather feedback for prioritizing additional connectors and tool integrations.