EvoX: Rethinking the Agent Harness as a Runtime Policy
Story 14, from arXiv
EvoX is a new agent runtime framework that improves the success rate of LLM agents by treating the harness as an operational policy, rather than just a tool wrapper. It outperforms existing harnesses across multiple models and benchmarks.
Key points
- EvoX is a Rust-based agent runtime that treats the harness as an operational policy.
- It uses three mechanisms: Action Realization, Execution Continuity, and Outcome Grounding.
- EvoX achieves a 56.2% success rate across 691 task-model pairs, outperforming other harnesses by 3.0 percentage points.
- Results are model-dependent, with notable gains on GLM-5.3 and Kimi K3.
- EvoX operates without memory, evolution, or extensions.
- Contrast is a framework for evaluating harnesses with fairness constraints and evidence-based adjudication.
Why it matters
As LLM agents become more integrated with tools and environments, their success depends not just on model capabilities but also on how they are controlled and guided at runtime. EvoX offers a new approach to harness design that could improve the reliability and effectiveness of multi-model agent systems.
Counterpoint
The results are model-dependent, suggesting that EvoX's effectiveness may vary across different foundation models and environments, limiting its universal applicability.
Sources
This summary was written by a language model and may contain errors. Read the original for the full story.