CHILL-Harness reduces token use and execution time for long-horizon AI agents

Modern large language model agents rely on harnesses to manage context, tools, verification, and execution control for long-horizon tasks. But current harnesses often use hand-crafted or globally fixed policies, leading to unnecessary computational overhead and sometimes reduced task success. To address this, a team of researchers introduces CHILL-Harness (Counterfactual Harness Intervention Learning for Long-Horizon Agents). The framework treats harness orchestration as a causal learning problem, enabling adaptive workflow adjustments. CHILL-Harness has two main components: causal intervention effect learning, which estimates the advantage of different workflows using confidence-weighted execution evidence, and advantage-realizing causal orchestration, which allocates counterfactual reasoning only when sufficient advantage is expected. It also includes success-preserving objectives and authorization constraints to ensure reliability.

In experiments spanning diverse long-horizon tasks—information seeking, software engineering, and terminal interaction—CHILL-Harness consistently preserved or improved task success while substantially reducing token consumption and execution time. The method outperformed static policies by adapting to task demands, execution environments, and evolving states. This work highlights a shift from rigid harness designs to intelligent, adaptive orchestration, promising more efficient and scalable LLM agents for complex, multi-step problems.