The History of AI

From answering to acting

Language models go from writing answers to using tools, browsers, and computers, and to working on their own for hours.

  1. Chain of thought (January 28, 2022)

    Chain of thought gets the model to reason in writing, but only with what it already knows.

  2. ReAct (October 6, 2022)

    ReAct interleaves that reasoning with actions, such as searching Wikipedia, and uses what it observes to decide the next step.

  3. SWE-bench (October 10, 2023)

    Meanwhile, SWE-bench proposes measuring models on real GitHub issues, much closer to actual work than the benchmarks of the time.

  4. Computer use in Claude (October 22, 2024)

    Anthropic applies the same loop to a computer. Claude looks at screenshots, moves the cursor, and types.

  5. Model Context Protocol (November 25, 2024)

    Meanwhile, Anthropic publishes a common protocol for connecting assistants to the systems where data lives.

  6. Operator, Deep Research, and Claude Code (January 23, 2025)

    Within a month, three specialized agents arrive. Operator browses the web, Deep Research investigates and writes reports, and Claude Code programs from the terminal.

  7. The time horizon of agents (March 18, 2025)

    METR proposes measuring agents by the length of the tasks they complete. That horizon doubles roughly every seven months, and even faster on SWE-bench Verified.

  8. ChatGPT agent (July 17, 2025)

    OpenAI combines Operator's browsing and Deep Research's investigation in a single agent with its own virtual computer and connectors to other applications.

  9. GPT-5.4 (March 5, 2026)

    Computer use stops being a separate product. GPT-5.4 has it built in and surpasses the human baseline on OSWorld-Verified.

Open in the interactive timeline