Across 691 common task–model pairs, the Rust-based EvoX runtime achieves a peak numerical success rate of 56.2%, surpassing the average of five alternative agent harnesses by approximately 3.0 percentage points. For business leaders deploying autonomous systems, this figure highlights a critical operational shift. High token spend and execution timeouts are frequently caused not by flawed reasoning in foundation models, but by structural execution failures at the software boundary where models interact with real-world tools.
Building reliable autonomous agents requires looking beyond the raw reasoning scale of foundation models to the deterministic controls governing their execution environment. The economic question for enterprise infrastructure is whether a runtime harness can systematically prevent execution drift and reduce compute waste without requiring expensive foundation model fine-tuning. Under a 20,480-token window and a 480-second timeout on 169 SWE-bench Verified tasks, wrapping a frozen model in an active closed-loop harness increased the mean per-task fail-to-pass fraction from 28% to 49% while boosting resolved tasks from 43 to 72. EvoX advances this paradigm by establishing an explicit division of operational logic.
The Architecture of The Second Policy
In a paper titled *EvoX: Rethinking the Agent Harness as a Runtime Policy*, researchers Haoyang Zhang, Qumeng Sun, Xiao Zhang, Xiang Liu, Zhidong Fan, and Junjie Wang from the Infinite Evolution Lab and EvoMap present EvoX as a multi-model agent runtime guided by what they call The Second Policy. This design framework separates an agent's workflow into two distinct operational layers: the Semantic Policy, embodied by the foundation model to interpret tasks and plan actions, and the Operational Policy, embodied by the deterministic Rust harness to control tool execution and manage state transitions.
"We present EvoX, a Rust-based, multi-model agent runtime guided by The Second Policy, which views the harness as an Operational Policy complementing the foundation model’s Semantic Policy."
Existing harness platforms prioritize distinct operational targets: SWE-agent emphasizes task-oriented agent–computer interfaces, OpenHands provides a general-purpose development environment, Claude Code and Codex support tool-driven coding workflows, and Pi adopts a minimal, extensible agent loop. EvoX replaces ad-hoc prompt wrapper scripts with three deterministic mechanisms: Action Realization, Execution Continuity, and Outcome Grounding. Action Realization detects malformed or text-leaked tool calls, validates tool schemas, and normalizes compatible arguments. Execution Continuity combines failure-recovery guidance, repeated-error detection, runtime circuit breakers, and context management to stop unproductive execution loops. Outcome Grounding forces models to align reported completion claims with observed environmental feedback.
Cross-Provider Benchmarks and Economic Metrics
To evaluate harnesses without suppressing native execution behaviors, the researchers introduced Contrast, a controlled framework integrating explicit fairness constraints, provider-bound request verification, and evidence-based adjudication that distinguishes locked, shared, and harness-native conditions. Comparative testing showed that EvoX led overall.
"Its larger advantages on GLM-5.3 (+10.7 points) and Kimi K3 (+7.9 points) highlight the model-dependent effectiveness of harness design and the potential for cross-provider adaptability."
These performance gains were achieved without incorporating long-term memory, evolution protocols, or external system extensions. By resolving schema errors and preventing infinite retry cycles at the runtime level, deterministic harnesses stop agents from exhausting token allocations on broken tool calls. Engineering teams deploying multi-model agent pipelines can audit their runtime architecture to verify whether tool schema validation and retry circuit breakers are handled by a compiled, deterministic harness layer rather than left to prompt instructions inside the prompt context.
For comparison
Back in late August an active closed-loop harness already demonstrated that tightly constraining an immutable model could boost stability without altering a single parameter. The EvoX framework turns this containment into an operational runtime in Rust, substituting ad-hoc prompt scripts with deterministic circuit breakers and API validation. Will the deterministic constraints of the EvoX Rust harness maintain its 56.2% success rate once workloads move past curated environments like SWE-bench Verified into noisy real-world APIs?
Our stories: Agent Harness Engineering Outperforms Model Upgrades (August 28).