EvoX: Rethinking the Agent Harness as a Runtime Policy
Summary
- Researchers introduce EvoX, a Rust-based, multi-model agent runtime that treats the agent harness as an operational policy complementing foundation models.
- EvoX uses Action Realization, Execution Continuity, and Outcome Grounding to provide deterministic runtime control and address tool-call mismatches.
- The authors also introduce Contrast, a controlled framework for evaluating harnesses using fairness constraints, request verification, and adjudication.
- Across five benchmarks, six foundation models, and six harnesses, EvoX achieved a 56.2% success rate, outperforming comparative harnesses.
Why it matters
EvoX demonstrates how treating agent harnesses as runtime operational policies can improve the reliability and execution success of large language model agents across multiple providers.