Agent-TARS

Run benchmarks with Agent-TARS

This page shows how to run various benchmarks using Agent-TARS.

Installation

npm install -g @agent-tars/cli@0.3.0

Basic Usage

uv run scripts/run.py \
  --agent Agent-TARS \
  --benchmark <benchmark_name> \
  --mode <run_mode> \
  [options]

Run LexBench-Browser

# Run first 5 tasks
uv run scripts/run.py \
  --agent Agent-TARS \
  --benchmark LexBench-Browser \
  --mode first_n --count 5

# Use no-login subset
uv run scripts/run.py \
  --agent Agent-TARS \
  --benchmark LexBench-Browser \
  --split no_login \
  --mode first_n --count 5

Run Online-Mind2Web

# Run first 5 tasks
uv run scripts/run.py \
  --agent Agent-TARS \
  --benchmark Online-Mind2Web \
  --mode first_n --count 5

# Use Hard30 subset
uv run scripts/run.py \
  --agent Agent-TARS \
  --benchmark Online-Mind2Web \
  --version 20251214 \
  --split Hard30 \
  --mode all

Run BrowseComp

uv run scripts/run.py \
  --agent Agent-TARS \
  --benchmark BrowseComp \
  --mode first_n --count 5

Common Parameters

Parameter	Description	Default
--mode	Run mode (all, first_n, sample_n, specific)	all
--count	Number of tasks	1
--split	Data subset	All
--timeout	Timeout in seconds	From config
--skip-completed	Skip completed tasks	False
--debug	Debug mode	False