Coding agents have enough intelligence, but they need better ways to specify, manage, reuse, and verify their work before people can trust their outcomes.
2
Recursive language models treat the prompt as an external object that agents explore through code execution and recursive calls to other models.
3
Claude Code Dynamic Workflows and OpenProse can turn coding agents into recursive systems that break down tasks, assign work, check results, and reuse successful sessions.
Summary
Raymond Weitekamp argues that coding agents are currently "mismanaged geniuses." Their raw intelligence is often sufficient, but they do not reliably deliver outcomes. He presents recursive language models as a way to combine reasoning with code execution. An RLM keeps the full prompt outside the model context, explores it symbolically through an executable environment, and delegates parts of the problem to recursive sub-agents. Weitekamp connects this design to coding agents, where recursion can support large codebase analysis, parallel refactoring, audits, adversarial review, and reusable workflows. He distinguishes agent-chosen decomposition from fixed map-reduce pipelines. He also discusses Claude Code Dynamic Workflows and OpenProse, a markdown-based programming language whose code is compiled by a coding agent. OpenProse can declare sub-agent responsibilities, required tools, and required skills, then turn a successful session into a reusable workflow. The talk is practical and opinionated about reliability: orchestration and verification matter more than simply increasing model intelligence.
Reliable outcomes require managing the model's work
Weitekamp says the main limitation of current agents is not intelligence. Models know a great deal, but they cannot reliably deliver outcomes, so users cannot trust them. He contrasts getting an almost complete SaaS app from one long prompt with a later incident in which Claude Code emptied the contents of his Solana wallet. He calls today's agents "mismanaged geniuses," borrowing the phrase from Alex Zeng, Zhi Li, and Omar Khattab at MIT. The missing layer is a way to specify, manage, reuse, and verify work.
An RLM computes over an external prompt through recursive calls
In a recursive language model, the context itself is the object of computation. The full prompt can be a file or many files rather than a short user query. An executable environment, originally a Python REPL in the paper, lets the model inspect and manipulate that external prompt symbolically instead of reading everything into its context window. The model can call other LLMs or RLMs to perform parts of the symbolic analysis, then combine those results into a final answer. The resulting structure is a tree of recursive calls.
RLMs combine code execution with reasoning at inference time
Weitekamp describes RLMs as a new approach to inference-time or test-time compute because they combine two existing ideas. Chain-of-thought prompting evolved into reasoning models with explicit reasoning tokens, while models also gained function calling and parallel tool calling. RLMs put reasoning and code execution together. The code execution is part of the reasoning process, and recursive calls let the system send a specific part of a problem to another model before returning to the parent.
Recursive execution can extend context and improve long reasoning
The talk gives several reported advantages of RLMs. They can process information many orders of magnitude larger than the model's context window, including tens of millions of tokens. Weitekamp says the default RLM harness also works as a strong memory system, with further improvements possible through modification. He cites results from the DSPy implementation on long reasoning tasks, where Qwen 3 0.59B running as an RLM beat larger frontier models on the Long CoT benchmark. These tasks require enough reasoning steps that ordinary reasoning models lose the thread.
An RLM needs model-directed decomposition, not just parallel calls
Weitekamp offers a rubric for deciding whether a system is an RLM. It needs an executable environment, an externalized prompt, code that calls the model, model-chosen decomposition into sub-calls, and symbolic state that remains outside the context. Plain LLMs and retrieval systems do not meet these conditions. Coding agents, sub-agents, and loops come close. A hardcoded map-reduce system such as lambda RLM also falls short in his view because the model does not decide how to decompose the problem. That decision is what makes the system agent-directed.
A recursive coding agent calls versions of its own harness
Weitekamp first wrapped Alex Zeng's RLM package as a command-line tool so a coding agent could send it a task such as searching a 100-million-token corpus. He then explored making the coding agent itself recursive, with the harness calling the same harness again. He called the first implementation Y Pi, using Pi, a minimal and extensible coding agent, as its base. After initially needing a fork, he says Pi's extension system evolved enough to support a pure recursive extension. The resulting setup can run Pi inside Pi inside Pi, with a configurable recursion depth.
Different runtimes can provide the recursive environment
The talk surveys several implementations. Alex's original RLM package and DSPy's pi.rlm are Weitekamp's preferred tools for benchmarking. AXE uses a TypeScript approach that lets one AXE agent write an interface to another AXE agent and recurse further. Dan's Unix RLM uses Bash and the Linux file system as its environment. OpenProse takes another route by letting a coding agent express recursive workflows in a language that can run with agents such as Claude Code or Codex.
OpenProse declares work, dependencies, and reusable workflows
OpenProse is a programming language written as a markdown specification in logical English. A coding agent compiles it, and a command such as prose write can generate a .prose.md file. The language can assign parts of a problem to sub-agents and require the parent session to verify their work. It can also declare skills and command-line tools as dependencies, so each sub-agent receives the capabilities needed for its assigned role. Another workflow can inspect a successful "golden session" from Claude Code, Codex, or Pi and turn it into a reusable Prose program.
"The intelligence is there, and the missing layer is how do we specify and manage and reuse and verify the work."02:00
Who should watch
You are building coding agents that produce inconsistent results and need explicit decomposition, verification, or reusable workflows.
Your tasks involve large files, large codebases, long reasoning chains, or work that can be split among specialized sub-agents.
You want to compare Claude Code Dynamic Workflows with a portable, markdown-based approach that can run across coding agents.
