fast-rlm
Recursive Language Models
fast-rlm is an inference technique where an LLM interacts with arbitrarily long prompts through an external REPL. The LLM can write code to explore, decompose, and transform the prompt. It can recursively invoke sub-agents to complete smaller subtasks. Crucially, sub-agent responses are not automatically loaded into the parent agent's context — they are returned as symbols or variables inside the parent's REPL.
How it works
User Query
|
v
Root Agent (primary_agent)
|-- writes Python code in REPL
|-- calls llm_query() to spawn sub-agents
| |
| v
| Sub-Agent (sub_agent)
| |-- explores a chunk of context
| |-- returns result as a variable
| v
| (result flows back as symbol, not raw text)
|
v
Final Answer
The root agent orchestrates the task by writing Python code. When it needs help with a subtask, it calls llm_query() which spawns a child agent. Child agents can spawn their own children, up to max_depth levels deep. Each agent has a budget of max_calls_per_subagent LLM calls.
Features
- Recursive decomposition — agents spawn sub-agents to handle subtasks
- REPL-based reasoning — agents write and execute Python code iteratively
- Budget controls — set hard limits on depth, calls, and dollar spend
- Any OpenAI-compatible API — works with OpenRouter, OpenAI, local models, etc.
- Structured logging — every step logged as JSONL, viewable in an interactive TUI
Learn More
Read the original RLM paper on arXiv
Note
fast-rlm is an independent implementation of the ideas from the RLM paper. This project is not affiliated with the original authors.
Support
If you find this helpful, consider supporting on Patreon — it hosts all code, projects, slides, and write-ups from the YouTube channel.