ClawGym-Bench: 200 Diagnostic Tasks for AI Agent Evaluation

200 diagnostic instances for Claw-style agents, each containing a user instruction, mock workspace resources, and a task-specific verifier. The benchmark was created by RUC-AIBOX and last updated on May 15, 2026. It uses a difficulty-aware filtering process for task selection.

Use Cases

• Benchmarking AI agent performance based on the 200 diagnostic tasks.
• Analyzing agent failure modes using the task-specific verifiers described.
• Comparing verification methods based on the 156 code-based and 44 hybrid verification tasks.
• Developing new agent training curricula based on difficulty-aware filtered tasks.

Strengths

• Contains 200 distinct diagnostic instances for agent evaluation.
• 156 tasks use code-based verification, providing objective scoring.
• 44 tasks use a hybrid verification method with a defined 0.7/0.3 weighting scheme.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

RUC-AIBOX

Collection Method

Selected through difficulty-aware filtering.

Time Range

null

Freshness

Last updated 2026-05-15 07:25:39; freshness should be verified.

Geography

null