{
  "video": {
    "id": "3hXJI2q0Jz8",
    "title": "Recursive Coding Agents - Raymond Weitekamp, OpenProse",
    "duration": 1428,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "Online"
  },
  "analysis": {
    "video_id": "3hXJI2q0Jz8",
    "title": "Recursive Coding Agents - Raymond Weitekamp, OpenProse",
    "one_liner": "Raymond Weitekamp argues that recursive language models — where the prompt is an externalized variable manipulated symbolically in a REPL and the model decides its own decomposition into sub-agents — are the next test-time-compute paradigm, and shows how to apply them to coding agents via pi extensions, Claude Code dynamic workflows and OpenProse.",
    "summary": "The talk's thesis is that \"today's agents are mismanaged geniuses\" — intelligence is no longer the bottleneck, reliable specification, orchestration and verification are. Weitekamp explains RLMs as a marriage of reasoning and code execution in which the context itself is the object of computation, cites results (tens of millions of tokens processed, top-10 memory system out of the box, SOTA on the Long CoT benchmark), and gives a five-point rubric for what actually counts as an RLM versus a hardcoded map-reduce. He then demos his own recursive coding agents — an RLM wrapped as a CLI tool call, Y-pi / pi-recursive where pi calls pi to arbitrary depth — and surveys Claude Code dynamic workflows and OpenProse as ways to turn any coding agent into an RLM.",
    "key_points": [
      "The motivating anecdote: one day a nearly fully working SaaS app from a single (long) prompt, the next day Claude Code emptied the entire contents of his Solana wallet — the gap is trust, not intelligence.",
      "\"Mismanaged genius\" framing comes from Alex Zeng, Z Li and Omar Khattab at MIT, authors of the original recursive language models paper (blog post October 2025).",
      "RLM definition: the full prompt is a variable (a file or many files) rather than a user query, the agent explores it symbolically in a REPL (Python in the original paper) and spawns sub-LLMs/sub-RLMs at whatever recursion depth you allow.",
      "Results cited: RLMs process information many orders of magnitude larger than their context window (tens of millions of tokens); the default unmodified RLM harness is already a top-10 memory system; and in the DSPy implementation on the new Long CoT benchmark, Qwen 3.5 9B as an RLM — runnable on a laptop — beats Opus and GPT-5.4 running as plain LLMs.",
      "\"Too hot to benchmark\": Symbolica's RLM harness Agentica scored 30-something percent on ARC-AGI-3 within hours of release when frontier models were at 2–3%, and the ARC Prize team responded with what he reads as a consolation tweet plus a refusal to run the full private evaluation; the Long CoT maintainers instead added a separate open-harness leaderboard.",
      "His RLM rubric (with a companion GitHub repo): executable environment, externalized prompt, code is the thing calling the model, the model picks the decomposition, state stays symbolic. Plain LLMs and RAG fail it; coding agents with sub-agents and loops get close; hardcoded map-reduces like lambda RLM fail because the model isn't choosing the decomposition.",
      "Experiments: first wrapping Alex's RLM package as a CLI so a coding agent gets an RLM as a tool call (e.g. to sift a 100M-token corpus), then Y-pi — named for the lambda-calculus Y combinator plus Mario's minimal, extension-oriented pi agent. A fork was originally required; pi extensions have since evolved so pi-recursive now works as a pure extension, pi calling pi calling pi at any depth.",
      "Omar Khattab now says Claude Code is finally an RLM thanks to dynamic workflows; Weitekamp wrote two Claude Code workflows to illustrate the line — a hardcoded map-reduce (not an RLM) and a deep-research-over-a-filesystem one (is) — and points to the \"A Harness for Every Task\" blog post with its six workflow patterns.",
      "OpenProse is a programming language compiled by your coding agent, not your computer: a markdown spec in logical English, written for you by a `prose write` command, that can explicitly declare sub-agent work, have the parent session verify it, and wire specific skills and CLI tools in as hard dependencies for each sub-agent.",
      "Newest OpenProse feature: take a \"golden session\" from Claude Code, Codex or pi and have a prose program deconstruct it into a reusable recursive workflow, so a good day becomes repeatable."
    ],
    "takeaways": [
      "Stop waiting for more raw intelligence — invest in the behavioral/orchestration layer: how work is specified, decomposed, verified and reused.",
      "Give your coding agent an RLM as a tool call for large-corpus work instead of stuffing context: wrap an RLM package as a CLI, or use pi-recursive/Y-pi to make the agent harness call itself at a chosen depth.",
      "Check your 'recursive' setup against the rubric — if you hardcoded the map-reduce, the model isn't choosing the decomposition and you don't have an RLM.",
      "If you want recursion without Claude Code dynamic workflows, write OpenProse .prose.md files and declare sub-agent roles with their required skills and tools as dependencies, so a sub-agent can't run under-equipped.",
      "Capture a session that went well as a prose program rather than hoping to re-prompt your way back to it."
    ],
    "topics": [
      "agents",
      "recursive-language-models",
      "coding-agents",
      "test-time-compute",
      "orchestration",
      "benchmarks",
      "context-engineering",
      "reliability"
    ],
    "tools": [
      "OpenProse",
      "DSPy",
      "Claude Code",
      "Codex SDK",
      "pi",
      "pi-recursive / Y-pi",
      "AXE",
      "Unix RLM",
      "Agentica",
      "Symbolica",
      "ARC-AGI-3",
      "ARC Prize",
      "Long CoT benchmark",
      "Qwen 3.5 9B",
      "Opus",
      "GPT-5.4",
      "lambda RLM",
      "Turing Post",
      "MIT",
      "recursivecodingagents.com"
    ],
    "quotes": [
      {
        "text": "The thesis is today's agents are mismanaged geniuses. The intelligence is there, and the missing layer is how do we specify and manage and reuse and verify the work.",
        "at": "02:00",
        "url": "https://www.youtube.com/watch?v=3hXJI2q0Jz8&t=120s"
      },
      {
        "text": "I like to say that in an RLM, the context itself is the object of computation.",
        "at": "02:47",
        "url": "https://www.youtube.com/watch?v=3hXJI2q0Jz8&t=167s"
      },
      {
        "text": "Qwen 3.59B as an RLM can beat Opus and GPT-5.4, all the top frontier models as LLMs on these long reasoning tasks.",
        "at": "06:55",
        "url": "https://www.youtube.com/watch?v=3hXJI2q0Jz8&t=415s"
      },
      {
        "text": "I don't care whether it's latent space or reasoning tokens or code execution. I want results. And I want AI programs that get those results.",
        "at": "08:57",
        "url": "https://www.youtube.com/watch?v=3hXJI2q0Jz8&t=537s"
      }
    ],
    "words": 4302
  },
  "summary_url": "/#3hXJI2q0Jz8",
  "transcript": {
    "html": "/transcripts/3hXJI2q0Jz8.html",
    "txt": "/transcripts/3hXJI2q0Jz8.txt",
    "vtt": "/transcripts/3hXJI2q0Jz8.vtt"
  }
}