CHILL-Harness Treats Agent Orchestration as a Causal Intervention Problem, Cutting Tokens and Wall Time Without Losing Task Success
CHILL-Harness (arXiv 2607.25825, 2026-07-28) targets the layer builders actually control — the harness coordinating context, tools, verification, and execution — arguing that today's harnesses rely on hand-crafted or globally fixed policies that mismatch task demands and burn compute. It formulates adaptive orchestration as causal learning: a causal intervention effect component estimates workflow advantage from confidence-weighted execution evidence, and an advantage-realizing orchestration component only applies adjustments backed by sufficient expected advantage, with a success-preserving objective and advantage-margin authorization constraints. Across heterogeneous long-horizon tasks spanning information seeking, software engineering, and terminal interaction, it preserves or improves task success while substantially reducing token consumption and execution time.
↳ Follow the thread