How microsoft/fara works
Fara is a visual computer-use agent: a user task enters through the CLI, the agent consults a model, acts on a browser, and records the run. The repository also includes WebEval, a separate evaluation and trajectory-verification package; the README identifies WebTailBench and the Universal Verifier as major capabilities. Some runtime modules are not sampled, so the graph omits uncertain implementation-level links and uses the README for documented workflows.
Main entry points: run_fara.py Fara CLI · fara15_agent.py Fara 1.5 agent · run_context.py Run context · websurfer.py Web surfer system
Components 20
Agent runtime
- Fara CLI
src/fara/run_fara.py - Fara 1.5 agent
src/fara/agents/fara/fara15_agent.py - Fara 7B agent
src/fara/fara_7b/fara_agent.py - Action prompts
src/fara/agents/fara/_prompts.py - Model client
src/fara/clients/wrapper.py
Browser interaction
- Browser environment
src/fara/environments/playwright/environment.py - Playwright controller
src/fara/environments/playwright/playwright_controller.py - 7B browser controller
src/fara/fara_7b/browser/playwright_controller.py
Run data
- Run context
src/fara/core/run_context.py - Trajectory data
src/fara/core/data_point.py
Evaluation and verification
- Evaluation core
webeval/src/webeval/core.py - System interface
webeval/src/webeval/basesystem.py - Web surfer system
webeval/src/webeval/systems/websurfer.py - Trajectory format
webeval/src/webeval/trajectory.py - Trajectory verifier CLI
webeval/scripts/verify_trajectories.py - Universal Verifier
webeval/src/webeval/rubric_agent/mm_rubric_agent.py - Evaluation LLM client
webeval/src/webeval/oai_clients/graceful_client.py
Other components
- Task user
- Model endpoint
- WebTailBench tasks
Connections 18
- Task user → Fara CLI: submits task
README.md - Fara CLI → Browser environment: initializes browser
src/fara/run_fara.py - Fara CLI → Fara 1.5 agent: runs agent
src/fara/run_fara.py - Fara CLI → Fara 7B agent: selects agent
src/fara/run_fara.py - Fara CLI → Run context: creates run
src/fara/run_fara.py - Fara 1.5 agent → Action prompts: uses prompts
src/fara/agents/fara/fara15_agent.py - Fara 1.5 agent → Model client: requests completion
src/fara/agents/fara/fara15_agent.py - Model client → Model endpoint: calls endpoint
README.md - Fara 1.5 agent → Run context: logs actions
src/fara/agents/fara/fara15_agent.py - Run context → Trajectory data: writes events
src/fara/core/run_context.py - Fara 1.5 agent → Browser environment: acts through browser
- Fara 7B agent → 7B browser controller: uses controller
src/fara/fara_7b/fara_agent.py - Evaluation core → System interface: uses interface
webeval/src/webeval/core.py - Web surfer system → System interface: implements interface
webeval/src/webeval/systems/websurfer.py - Web surfer system → Trajectory format: records trajectory
webeval/src/webeval/systems/websurfer.py - Trajectory verifier CLI → Universal Verifier: runs verifier
webeval/scripts/verify_trajectories.py - Universal Verifier → Evaluation LLM client: requests judgments
webeval/src/webeval/rubric_agent/mm_rubric_agent.py - WebTailBench tasks → Trajectory verifier CLI: supplies tasks
README.md