| name | uniclawbench-proactive-agents-real-world-tasks |
| description | Capability-driven benchmark for evaluating proactive agents in dynamic real-world settings. UniClawBench evaluates five foundational capabilities (Skill Usage, Exploration, Long-Context Reasoning, Multimodal Understanding, Cross-Platform Coordination) across 400 bilingual tasks in live Docker containers with closed-loop evaluation. Activation: proactive agents, real-world benchmark, agent evaluation, capability-driven, multimodal agents, closed-loop evaluation. |
| metadata | {"arxiv_id":"2607.08768","published":"2026-07-09","authors":"Zhekai Chen, Chengqi Duan, Kaiyue Sun, Bohao Li, Yuqing Wang","tags":["proactive-agents","real-world-benchmark","agent-evaluation","capability-driven","multimodal-agents"]} |
UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks
Overview
The rapid development of large language models and multimodal large language models has accelerated the emergence of proactive agents capable of operating everyday tools and assisting users in real-world environments. UniClawBench is the first capability-driven benchmark designed to evaluate proactive agents in dynamic, real-world settings, addressing limitations of existing benchmarks that rely on sandboxed environments and single-turn evaluation paradigms.
Key Innovations
Capability-Driven Task Taxonomy
- Five foundational model capabilities: Skill Usage, Exploration, Long-Context Reasoning, Multimodal Understanding, and Cross-Platform Coordination
- Disentangles model capabilities from framework-level design choices
- Enables identification of root causes of agent failures
Live Docker Container Evaluation
- Tasks executed in live Docker containers rather than static pre-recorded answers
- Fine-grained, step-by-step completion checkpoints
- Realistic multi-turn human feedback simulation
Closed-Loop Evaluation Strategy
- Executor agent performs tasks
- Hidden supervisor agent evaluates
- User agent simulates human feedback
- Prevents grading criteria leakage
Methodology
- Capability Taxonomy: Design 400 bilingual tasks across five capability dimensions
- Live Environment: Deploy tasks in Docker containers with real system state
- Closed-Loop: Multi-agent evaluation simulating real user interaction
- Framework Comparison: Evaluate state-of-the-art models under multiple agent frameworks
Implications
- Framework design and base model capabilities jointly shape real-world performance
- Capability-driven benchmarks enable targeted improvement of agent systems
- Live evaluation reveals failure modes invisible to static benchmarks
- Supports the growing field of proactive agent deployment
Pitfalls
- Docker container setup may limit scalability for very large evaluations
- Bilingual tasks may introduce language-specific biases
- Framework comparison depends on framework maturity and optimization
- Closed-loop evaluation complexity may increase evaluation cost
Activation Keywords
proactive agents, real-world benchmark, agent evaluation, capability-driven, multimodal agents, closed-loop evaluation, Docker benchmark, agent frameworks, UniClawBench
Paper Reference
arXiv:2607.08768 - "UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks" (Jul 2026)