Agent Tool-Use Trajectories with Process-Level Labels

AgentProcessBench provides 1,000 full agent trajectories for process-level evaluation of tool-using agents. The benchmark was created by LulaCola and published in March 2026. Each trajectory includes multi-turn messages, tool definitions, tool-use traces, reference outputs, and step-wise process labels.

Use Cases

• Evaluate agent reasoning by analyzing step-wise process labels against tool-use traces.
• Benchmark tool-calling accuracy using tool definitions and multi-turn message sequences.
• Train process-supervised models on trajectories containing reference outputs and intermediate labels.
• Analyze failure modes in agent planning by comparing trajectories from the bfcl, gaia_dev, hotpotqa, and tau2 subsets.

Strengths

• Contains 1,000 complete agent trajectories for evaluation.
• Includes step-wise process labels for detailed analysis of agent reasoning.
• Balanced with 250 examples from each of four distinct subsets (bfcl, gaia_dev, hotpotqa, tau2).

Limitations

• Limited to 1,000 total examples, which may be insufficient for training large models.
• Potential bias towards the specific tasks and domains represented by the four source subsets.
• Column structure and specific data types are not documented.

Provenance

Source

LulaCola via Hugging Face.

Collection Method

Curated benchmark for process-level evaluation, sourced from four subsets.

Freshness

Last updated March 2026.