← back
Recursive Coding Agents - Raymond Weitekamp, OpenProse
Takeaway
Reliable agent outcomes depend on how work and context are decomposed and verified, not only on model intelligence.
Summary
- Recursive language models externalize a large prompt as data and use a REPL to inspect it symbolically rather than loading everything into context.
- The model chooses how to decompose work into recursive calls while code maintains intermediate state.
- Weitekamp reports strong memory and long-reasoning results from RLM harnesses, including smaller models outperforming larger models under different harness conditions.
- Recursive coding agents extend this approach to agents and subagents, emphasizing specification, management, reuse, and verification of work.
recursive-language-modelscoding-agentscontext-management
Original description
Recursive Language Models (RLMs) represent a powerful new paradigm of inference-time compute. We discuss many different ways to apply the principles of RLMs to coding agents, towards higher performance and reliability. We briefly define RLMs, showcase many of their performance advantages, then share how the RLM paradigm can be mapped onto coding agents. We strive to settle the long-standing debate, "Isn't Claude Code with sub-agents a RLM?" Finally, we showcase how both Claude Code Dynamic Workflows and OpenProse can specifically guide coding agents to recursively solve complex tasks with declarable outcomes. The "slides" for this presentation are available as an interactive website at https://recursivecodingagents.com and a companion repo with specific code examples is available at https://github.com/rawwerks/recursive... Speakers: Raymond Weitekamp (OpenProse): Dr. Raymond Weitekamp is a PhD chemist, serial entrepreneur, artist, independent researcher, and AI Engineer at OpenProse. X/Twitter: https://x.com/raw_works LinkedIn: / raymondweitekamp GitHub: https://github.com/rawwerks