Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2607.25825.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

CHILL-Harness is a method for adaptive orchestration in LLM agent harnesses that uses counterfactual causal learning. It estimates workflow advantage from confidence-weighted execution evidence and selectively applies only adjustments with sufficient expected benefit, reducing token use and execution time while preserving task success across long-horizon tasks.

Frames harness adaptation as causal intervention rather than fixed policy, with reported token and time reductions on long-horizon benchmarks. Relevant if you build or tune agent systems.

Cited by

No citations on record.