Recursive Agent Harnesses

The article introduces the Recursive Agent Harness (RAH), a framework that enhances long-context reasoning by employing a recursive architecture where a parent agent generates and executes scripts that spawn subagent harnesses for parallel processing. Evaluations show that RAH significantly improves performance on coding tasks, increasing the Codex baseline from 71.75% to 81.36% on the Oolong-Synthetic benchmark, with further enhancements to 89.77% using Claude Sonnet 4.5 as the backbone. This advancement is crucial for practitioners as it demonstrates a scalable approach to complex task execution and fine-grained workload management in AI systems.