Recursive language models, explained (the idea under Claude "ultracode")

On May 28th, 2026, Anthropic shipped "dynamic workflows" in Claude Code, switched on by a setting that sounds like a cheat code: ultracode. Flip it, hand the AI a big job, and it does something strange — it stops trying to answer you. Instead it writes a program, and that program spins up an army of copies of the AI: fifty, sometimes a few hundred, each on its own slice of the work, while the one in charge runs the whole operation from code and never even holds most of what comes back.

People can't agree whether it's the biggest leap in a year or an expensive way to burn money — but that argument isn't the interesting part. This video is about how it pulls the trick off at all: the MIT idea underneath it, recursive language models — stop feeding the model your data, leave it in a variable, and make it write code to reach in for only the scraps it needs. We walk through why one overloaded context window "rots," the company-org-chart shape that fixes it (recursion is just a manager who's allowed to hire), the benchmark where the same models jump from basically zero to 76, why everyone's so careful with it (burned budgets, agents that hallucinate and then confirm each other), how one lab trained the skill into the model instead of prompting it, and when this is actually worth using. It ends on the shift it forces: you stop asking "how do I solve this?" and start asking "how would I structure a team to solve it?" If you want more under-the-hood breakdowns like this, subscribe — and tell me in the comments where you'd point an AI army first.

## Chapters
00:00 One AI writes a program that spawns an army of itself
00:50 "Isn't this just subagents?" — the one real difference
02:05 Recursive language models: stop feeding the model your data
02:58 Why one overloaded context rots — and the fix
05:01 The proof: same models, 0 → 76
06:18 It's everywhere — and why everyone's careful
07:43 Teaching the skill, and when to actually use it
09:12 The bigger shift: you draw the org chart

## Sources & further reading
- Anthropic — Dynamic Workflows in Claude Code (announcement) — https://claude.com/blog/introducing-dynamic-workflows-in-claude-code
- Anthropic — A harness for every task: Dynamic Workflows — https://claude.com/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code
- Recursive Language Models — Alex Zhang's blog (with interactive trajectories) — https://alexzhang13.github.io/blog/2025/rlm/
- Recursive Language Models (Zhang, Kraska, Khattab — MIT CSAIL), arXiv — https://arxiv.org/abs/2512.24601
- CodeWhale — the DeepSeek coding agent that wired RLM in, in the open — https://github.com/Hmbown/CodeWhale
- Kimi agent swarm + PARL (Moonshot), arXiv — https://arxiv.org/abs/2602.02276

Видео Recursive language models, explained (the idea under Claude "ultracode") канала Squintist