Computer Use Agent

Single-pass A2A agent for the CAR-bench track. Uses a reasoning-capable LLM (default: openai/gpt-5-mini with reasoning_effort=medium) plus a compact, domain-agnostic prompt consisting of six general agent rules. No hardcoded policy content, tool names, or task-specific lookup tables — all instructions come from the green agent at runtime.

The green agent assesses OSWorld desktop tasks, which are real-world Ubuntu Linux computer interaction issues. Creating and editing files, visiting websites, utilizing programs (such as LibreOffice, GIMP, and VLC), and adjusting system settings are examples of these tasks. By comparing the actual VM state (files, webpages, and app states) with the anticipated outcomes, the agent determines whether the task was successfully finished and returns a score ranging from 0.0 (failed) to 1.0 (succeeded). There are 385 tasks in the benchmark (370 original + 15 new) that assess actual desktop workflows across 10 domains: Chrome, GIMP, LibreOffice Calc/Impress/Writer, Multi-Apps, OS, Thunderbird, VLC, and VS Code.

Test dummy agent pushing

Planner execute verify agent Planner model create a list of intermediate goals, then ReAct agent execute actions to achieve this goal, when finish - the planner verify actions with summarized trajectory, after that

Try purple agent