Recursive Coding Agents - Raymond Weitekamp, OpenProse

TL;DR

• 1

  RLMs as the new test-time compute paradigm: By marrying chain-of-thought reasoning with code execution, RLMs allow models to symbolically explore massive contexts and decompose problems recursively.

• 2

  Small models beat giants with RLMs: Qwen 3.5 9B using an RLM harness outperformed GPT-5.4 and Opus on the Long CoT benchmark, running on a laptop.

• 3

  RLMs stir benchmark drama: The Symbolica team achieved 30%+ on Arc AGI-3 within hours using an RLM harness, prompting the Arc Prize team to dismiss the results and refuse private evaluation.

• 4

  Recursive coding agents apply RLM principles to agent frameworks: By making coding agents call themselves recursively (e.g., Y pie), they can self-decompose tasks and verify sub-agent outputs.

• 5

  Claude Code's dynamic workflows now make it an RLM: Anthropic's addition of dynamic workflows turned Cloud Code into a recursive agent system, settling earlier debates.

• 6

  OpenProse turns any coding agent into an RLM: Using a markdown-based language, developers can declare sub-agent workflows, dependencies, and verification steps, capturing golden sessions for reuse.