tinyhumansai/tinyagents

TinyAgents is a Rust toolkit for building AI agents, built around an idea called a recursive language model, or RLM. Most tools for building AI agents try to stuff everything the model needs into one growing block of text and hope the model can handle it. TinyAgents takes a different approach: instead of loading all the information at once, a long piece of text is treated like a small environment the model can explore step by step, breaking it into pieces and calling itself, or smaller helper models, on those pieces. The goal is to let an AI system work with far more information than would normally fit in its memory at one time, while keeping the process organized and inspectable. The project is inspired by a 2025 research paper on recursive language models from MIT, and it brings that idea into a working Rust system rather than just following the research results exactly. In TinyAgents, one AI agent can call another agent as if it were a simple tool, so a chain of agents working together looks like a series of function calls. It also includes a graph system, similar in spirit to tools like LangGraph, where each step in a workflow is a node that can pass control to the next step, pause, resume later, or even contain another whole workflow inside it. TinyAgents provides two small custom languages for describing this behavior: one is a fixed blueprint format for laying out a workflow in advance, and the other is an interactive command style where the AI model can write, run, and inspect small programs on the fly. The system also tracks how deep the calls go, keeps a record of costs and usage across all these nested calls, and can pause and resume long-running processes without losing progress. It works with multiple AI model providers rather than locking a user into just one. The README says the project does not try to reproduce the exact benchmark numbers from the original research paper. It is licensed under GPLv3, meaning any software built using it and distributed to others must also share its source code under the same license.