Back to directory

Agent S

An open-source computer-use agent that operates a real graphical interface, mouse, keyboard and screen, the way a person would, with successive versions reporting new state-of-the-art results on OSWorld.

Agent S

Agent S is an open-source computer-use agent that operates a real graphical interface, using the mouse, keyboard and screen the way a person does, rather than through an API built for automation. It has been developed across several published versions, from the original Agent S paper through Agent S2 and Agent S3 to a production agent called Sai, with each version's paper accepted at a machine-learning venue such as ICLR, COLM or TMLR.

Successive versions report new state-of-the-art results on OSWorld, a benchmark for computer-use agents: Agent S3 is reported as the first agent to exceed average human performance on OSWorld, and the production Sai agent is reported to top OSWorld 2.0 while costing less than comparable systems such as GPT- and Claude-based agents. Earlier versions also reported outperforming OpenAI's own computer-use agent and Anthropic's Claude Computer Use on the original OSWorld benchmark.

It is distributed as the gui-agents Python library, supporting macOS, Linux and Windows, with a hosted Simular Cloud offering to try it without local setup.

More in Agent frameworks

See all 902
  • Recurse

    recurse.run

    A serverless harness that lets a coding agent design, test, and deploy its own specialist sub-agents as tools, MCP servers, or bots, defined by a schema-validated manifest.

  • ProofForge

    shulgin.is-a.dev

    An AI-agent pipeline that decomposes a mathematics problem, proves each piece, and formalizes the result in Lean 4 and Mathlib until the Lean kernel accepts it, with six of its proofs already merged into Google DeepMind's formal-conjectures repository.

  • local-coder

    github.com

    Builds a local coding-agent team on top of Ollama and OpenCode by detecting the machine's hardware, recommending models for seven specialist roles, and testing whether each assigned model can actually use tools, edit a repository and pass verification.