LiveClawBench: Benchmark Trajectories for LLM Agents on Real-World Assistant Tasks

A pilot benchmark of 30 manually created tasks for evaluating LLM agents, created by Mosi-AI and last updated in April 2026. The dataset focuses on complex, real-world assistant tasks such as booking flights and debugging code. It introduces a Triple-Axis Complexity Framework to address gaps in existing benchmarks.

Use Cases

• Benchmarking agent performance on real-world tasks like flight booking based on the described task types
• Evaluating agent robustness across multiple difficulty axes based on the Triple-Axis Complexity Framework
• Training or fine-tuning agents for knowledge base curation based on the described task scope
• Analyzing agent failure modes in complex, integrated scenarios based on the benchmark's design goal

Strengths

• Introduces a Triple-Axis Complexity Framework for more holistic evaluation
• Focuses on 30 manually created, complex real-world tasks
• Last updated on 2026-04-08, suggesting recent maintenance

Limitations

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

Provenance

Source

Mosi-AI via Hugging Face

Collection Method

Manually created benchmark tasks

Time Range

Creation and update timeframe includes 2026

Freshness

Last updated 2026-04-08 11:10:10

Geography

null