Recursive Agent Harnesses
Elias Lumer, Sahil Sen, Kevin Paul, Vamse Kumar Subbiah
Recursive Agent Harnesses (RAH) extends the idea of recursive model calls โ where a language model repeatedly calls itself to handle long contexts โ to full agent harnesses with filesystem access, code execution, and planning tools. A parent agent writes and runs a script that spawns parallel subagent harnesses for compute-heavy subtasks and uses structured function calls for lighter ones. On Oolong-Synthetic, a long-context reasoning benchmark spanning up to 4M token contexts, RAH pushes GPT-5-backed Codex from 71.75% to 81.36%, with Claude Sonnet 4.5 as backbone reaching 89.77%. The gain is attributed to the harness architecture, not model differences.
No production traction yet โ zero citations and the GitHub references are arxiv aggregator lists, not implementation forks. The pattern itself is real and already visible in Anthropic's dynamic workflows and coding agents that spawn subagents, but this paper is naming and measuring it rather than shipping tooling. Builders working on orchestration layers or long-context agent pipelines should watch it as a framing reference, not a library to depend on.
Recursive language models (RLMs) showed that recursion over model calls is an effective strategy for long-context reasoning, and production coding agents have begun to write code that spawns subagents at scale, most recently in Anthropic's dynamic workflows. We name and study the pattern between these two lines of work, where the recursive unit is a full agent harness with filesystem tools, code execution, and planning rather than a model call with no tools. We call this the Recursive Agent Harness (RAH) and frame it as harness recursion, the code-first extension to the model recursion of RLMs. A parent agent generates and runs an executable script that spawns subagent harnesses in parallel for fine-grained workloads and uses structured function calls for small subtasks. We provide a controlled evaluation on long-context reasoning. With the backbone held fixed at GPT-5 to match the published Codex and RLM baselines, RAH improves the Codex coding-agent baseline from 71.75% to 81.36% on Oolong-Synthetic (199 samples, 13 context-length buckets up to 4M tokens), a gain attributable to the harness rather than the model. With a stronger backbone, Claude Sonnet 4.5, the same design reaches 89.77%.