Agent-S

Facts

Openness

Public code, licence not stated

Language

Python

Pinned version

2cb57fb (2025-12-16)

Topics checked

28 of 33

Kind of agent

Browser/GUI

Source repository (opens another site)

What it does notably well

• Budgets

  In gui_agents/s3 the CodeAgent tool has its own step budget, default 20 (code_agent_budget in OSWorldACI), separate from the CLI's 15-iteration loop; after it ends (done, fail or…

• Code mode

  Agent-S s3 runs model-written code through a separate CodeAgent (own LMMAgent, default budget 20) and returns to the worker a result dict with a model-written summary that the…

• Eval harness

  Agent-S's Behavior Best-of-N is an offline step after several OSWorld rollouts: BehaviorNarrator writes a fact caption for each step's before/after screenshots and…

  Agent-S evaluates on the external OSWorld benchmark: README.md points to osworld_setup/s3/OSWorld.md, which tells the user to copy the provided run files into an OSWorld checkout.

• Perception and grounding

  Agent-S s3 splits planning from grounding: the main model emits natural-language element descriptions, a separately configured grounding model (README recommends UI-TARS-1.5-7B)…

  Agent-S s3 turns the model's last code block, a call such as agent.click("description"), into a desktop action by eval() against the grounding agent in create_pyautogui_code, and…