Recursive Coding Agents - Raymond Weitekamp, OpenProse

Key Concepts

• Recursive Language Models (RLMs): A paradigm where the context is an object of computation, combining reasoning with symbolic code execution.
• Mismanaged Genius: The theory that current AI models possess sufficient intelligence but lack the orchestration, specification, and verification layers to deliver reliable outcomes.
• Test-Time/Inference-Time Compute: The shift from relying solely on pre-trained model weights to using active, recursive computation (reasoning + tool use) to solve complex problems.
• REPL (Read-Evaluate-Print Loop): The core loop where an agent interacts with an environment (e.g., Python, Bash) to perform symbolic manipulation.
• OpenProse: A markdown-based specification language that allows users to define workflows, dependencies, and sub-agent tasks, effectively turning any coding agent into an RLM.
• Y-Combinator (Y-Pie): A framework for creating fully recursive coding agents where an agent can call instances of itself to arbitrary depths.

---

1. The Problem: Reliability and "Mismanaged Geniuses"

Raymond Whitcomb argues that the primary bottleneck for AI agents is not a lack of intelligence, but a lack of reliability. He illustrates this with a personal anecdote: one day an agent successfully built a SaaS app, and the next day it emptied his crypto wallet. The thesis is that we have "mismanaged geniuses"—models that know everything but lack the framework to manage, reuse, and verify their work.

2. Recursive Language Models (RLMs)

RLMs represent a marriage of reasoning and code execution.

• Mechanism: Instead of loading everything into a context window, the model operates symbolically on the prompt (which acts as a variable).
• Recursion: The model decomposes a problem into sub-tasks, which are then handled by sub-agents (or sub-RLMs), working back up to a final answer.
• Performance: RLMs can process information orders of magnitude larger than standard context windows (tens of millions of tokens).
• Benchmark Success: Using the Long CoT (Chain of Thought) benchmark, smaller models (e.g., Qwen 3.59B) running as RLMs outperformed frontier models like GPT-5.4 and Opus because they could maintain the reasoning thread through recursive code execution.

3. Frameworks and Implementations

Whitcomb highlights several key projects that enable recursive coding:

• Agentica (Symbolica): A high-profile harness that achieved ~30% on the ARC-AGI benchmark (compared to 2-3% for standard models) by using RLM frameworks.
• DSPy: Used by the author to achieve state-of-the-art results on long reasoning tasks.
• AXE: A TypeScript-based variation of DSPy that is "agent-native," allowing agents to write interfaces for other agents.
• Unix RLM: A project by Dan at OpenProse that uses the Linux file system as the environment for the REPL.
• Claude Code (Dynamic Workflows): Recently updated to support dynamic workflows, allowing it to function as an RLM by enabling recursive, multi-step task delegation.

4. OpenProse: A Declarative Approach

OpenProse is presented as a solution for users who want to standardize agent behavior without learning complex syntax.

• Functionality: It uses a markdown-based "Prose" file to define the logic of an agent.
• Dependency Management: Users can explicitly declare skills and tools as dependencies for specific sub-agents, ensuring that an agent has the necessary environment to perform its assigned task.
• Golden Session Capture: A unique feature where an agent can deconstruct a successful "golden session" of work and turn it into a reusable, recursive Prose workflow.

5. Key Arguments and Perspectives

• The "Rubric" for RLMs: To be considered an RLM, a system must have:
  1. An executable environment.
  2. An externalized prompt.
  3. Code that calls the model.
  4. The model's ability to decompose problems into sub-calls.
  5. Symbolic state maintenance.
• The "No-Tool" Controversy: There is tension between RLM researchers and leaderboard maintainers. Some benchmarks (like ARC-AGI) prohibit tool calling, while researchers argue that RLMs are the only way to solve these problems effectively. Whitcomb advocates for separate "open harness" leaderboards to allow for innovation.

6. Synthesis and Conclusion

The transition from standard LLMs to recursive coding agents is the next evolution in AI. By shifting focus from raw model intelligence to orchestration and behavioral management, developers can create reliable, repeatable systems.

Main Takeaways:

1. Trust is Reliability: Reliability is achieved through structured, recursive workflows rather than just larger models.
2. Unified Reasoning: Tool calling and reasoning are not separate; they are the same process when executed recursively.
3. Actionable Insight: Developers should move toward defining workflows (via OpenProse or dynamic workflows) that allow agents to decompose tasks, verify sub-agent work, and maintain symbolic state.

Quote: "The intelligence is there, and the missing layer is how do we specify and manage and reuse and verify the work." — Raymond Whitcomb