Public talk website/deck, RLM rubric, and calibration examples for Raymond Weitekamp's Recursive Coding Agents talk at AI Engineer World's Fair 2026.
The live presentation is a website: https://recursivecodingagents.com
Modern coding agents are powerful, but reliability is still the bottleneck. This project explains how to apply the principles of Recursive Language Models (RLMs) to coding agents. RLMs are systems that keep context as symbolic state, use code to inspect and slice it, call models or agents over those slices, and aggregate the result through a verifiable process.
The repo is both a companion artifact for the talk and a public calibration set: it gives humans and agents concrete examples of run shapes that satisfy the RLM rubric, nearby shapes that do not, and how to judge the difference.
| If you want to... | Open |
|---|---|
| See the presentation website (slides) | https://recursivecodingagents.com |
| Understand the RLM definition | rlm-rubric/README.md |
| Apply the full seven-gate rubric | rlm-rubric/rlm-rubric.md |
| Judge a system with evidence | rlm-rubric/rlm-judging-methodology.md |
| Compare Claude Code workflow examples | claude-dynamic-workflows/ |
| Compare OpenProse examples | openprose/ |
recursive-coding-agents/
├── web/ public SvelteKit deck/site source
├── rlm-rubric/ RLM definition, seven gates, and judging method
├── claude-dynamic-workflows/ Claude Code workflow examples: RLM vs not-RLM
└── openprose/ OpenProse program examples: RLM vs not-RLM
The example folders are intentionally verdict-shaped. In claude-dynamic-workflows/ and openprose/, files under rlm/ are passing or intended-passing examples; files under not-rlm/ are nearby negative controls.
Read the verdicts by run shape, not by product label. A coding agent with subagents, loops, bash, durable files, or a repo handle should get credit for the gates it actually satisfies. It still is not a full RLM unless the run also externalizes the task context behind handles, lets the model choose a decomposition, makes programmatic model/subagent calls over constructed slices or subproblems, and aggregates symbolic intermediate state into the final answer.
An RLM moves the task context into a persistent executable environment as symbolic state, lets the root model work through handles and metadata, lets the model write code that inspects and slices that state, recursively calls models or agents over the slices, and returns the final answer through the outer model-call interface.
Copyright (c) 2026 Raymond Weitekamp (https://RAW.works). All rights reserved.