Coding agents aren't short on intelligence, they're short on reliability. Recursive decomposition turns the lucky golden session into a repeatable result.
Recursive Coding Agents - Raymond Weitekamp, OpenProse · Raymond Weitekamp
24 min total·Actually worth watching closely: ~7 min·2 must-watch clips
- 0:00 – 3:10Listen
Mismanaged geniuses
Opens by naming the problem: coding agents aren't insufficiently smart, they're unreliable. The same tool writes an almost fully working SaaS app one day, and empties the speaker's own Solana wallet the next.
Trust is really about controlled variance, not peak capability. What's missing is the orchestration layer, not raw intelligence.
The speaker is recounting his own war stories the whole way through and the screen barely moves off the title slide. Earbuds on the commute are plenty.▶ Jump to 0:00Speaker · Raymond Weitekamp - 3:10 – 7:04Listen
Turning context into something you can operate on
Explains the core mechanism of recursive language models: the full prompt is externalized into a variable, the agent manipulates it symbolically inside an executable environment, and it decides for itself how to carve the work into subcalls and work back up to an answer.
There's only one test that matters: the model must pick the decomposition itself. A hard-coded mapreduce doesn't count.
This stretch is conceptual derivation, built up verbally layer by layer. The slide adds nothing extra, so just follow the reasoning.▶ Jump to 3:10Speaker · Raymond Weitekamp - 7:04 – 11:00Skim
A small model beating the frontier
Puts up a run of benchmark results: information many orders of magnitude beyond the context window, a default harness that lands near the top ten as a memory system with no modifications, and Qwen 3.5-9B beating Opus-class models on long reasoning tasks.
What stalls frontier models on these problems is that the reasoning runs too many steps to hold the thread, not that any single step is too hard.
The screen shows a run of comparison tables and leaderboard screenshots. Glancing at the orders of magnitude and the model names is enough; reading line by line isn't worth it.▶ Jump to 7:04Speaker · Raymond Weitekamp - 11:00 – 14:06Listen
Upsetting the benchmark world
Covers the fallout: Symbolica's harness hit 30-something percent on Arc AGI 3 within hours of release (frontier models were at 2-3 percent) and was refused the full private evaluation by the Arc Prize team, while the long-reasoning leaderboard maintainers were pushed into opening a separate open-harness board.
Existing benchmarks assume no tool calling, which is exactly the line recursive agents cross. The better the score, the harder it is to get recognized.
This part is the community backstory and where each side stands, carried by narration with only a few scattered screenshots. Listening keeps you with it.▶ Jump to 11:00Speaker · Raymond Weitekamp - 14:06 – 17:30Watch
Coding agents crossing the line
Surveys the implementations out there (DSP.rlm, YPI, Axe, the Unix RLM) and explains how Claude Code went from not counting to arguably counting once dynamic workflows shipped.
Skip the definition fight. The question to ask is how to apply recursive decomposition to the agent in front of you.
The speaker walks through the projects on screen as he talks, and points to the public reply from the paper's author on the right. Something is being pointed at whenever he says "over here," so not watching costs you half a sentence.▶ Jump to 14:06Speaker · Raymond Weitekamp - 17:30 – 21:07Watch
How OpenProse is used
Introduces OpenProse: a language compiled by your coding agent rather than a compiler, written in markdown and logical English with no syntax to learn. It can explicitly declare what each subagent does, verify the subagents' output back in the parent session, and wire skills and CLI tools to specific subagents.
Any agent with a file system and subagents becomes a recursive agent once you write the spec, with its capability setup written into the contract too.
How the spec file is actually written and how the subagents are wired are both on screen, and this kind of structure is worth one look over three explanations. Keep the video in view.▶ Jump to 17:30Speaker · Raymond Weitekamp - 21:07 – 23:46Listen
Locking in the day it went brilliantly
Closes with where this lands: repo-scale migrations refactored in parallel then merged, recursive deep research, audits and bug sweeps, red-team style adversarial improvement, with the weight on golden-session reuse.
Deconstruct a session that went brilliantly into a recursive process you can rerun, and luck becomes a deliverable.
The ending lists scenarios and pulls the argument together. Steady pace, nothing on screen you need to watch, and you leave with the conclusion.▶ Jump to 21:07Speaker · Raymond Weitekamp