Skip to content

Home

Jump to bottom
Steven Enamakel edited this page Jun 29, 2026 · 3 revisions

TinyAgents Wiki

TinyAgents is a recursive language-model (RLM) harness for Rust: a typed, durable runtime where language models call models, agents call agents, graphs run graphs, and a model can author, compile, and run the very workflow it is standing inside — all as inspectable, checkpointed, policy-checked Rust.

The repository README.md is the marketing-level overview (hero, diagrams, 30-second start). This wiki is the implementation-oriented companion: how each surface works, how to set things up, how to develop against the crate, and the RLM execution model in depth.

Why "recursive"

TinyAgents is architected around the Recursive Language Model execution model: treat a long prompt as an external environment a model interacts with through a REPL — examining, decomposing, and recursively calling itself (or sub-models) over snippets — instead of stuffing everything into one context window. The crate brings that model to Rust as a typed, durable harness. Recursion shows up concretely as:

  • Sub-agents — a harness agent is exposed as a tool to another agent (harness::subagent), so orchestration is just a model calling a model.
  • Recursion policy + depth tracking — runs carry root_run_id / parent_run_id, a recursion limit is enforced, and child runs roll their events, usage, and cost up to the parent.
  • Graphs that run graphs — subgraph nodes (graph::subgraph) embed another compiled graph; the REPL can drive a graph from inside a graph node.
  • The REPL as the RLM core — .ragsh makes context and prompts runtime values; the model writes small programs, inspects output, and calls sub-models / sub-agents / sub-graphs as functions.
  • Self-authoring — a model can emit a .rag blueprint that compiles through the same registry-bound path as a human-authored file and runs on the same runtime the model is already executing in.

See Recursion and RLM for the full treatment and research lineage (Zhang, Kraska, Khattab, "Recursive Language Models," 2025).

The five surfaces

TinyAgents is organized around five surfaces, used consistently across the docs:

  1. Harness (src/harness/) — provider-neutral model calls, typed tools, middleware, structured output, streaming, usage/cost, retry/limits, cache, memory/embeddings, sub-agents, steering, summarization, testkit.
  2. Graph runtime (src/graph/) — durable typed state graphs: START/END, nodes, edges, conditional routing, commands, Send fanout, reducers/channels, checkpoints, interrupts, subgraphs, streaming, topology export, time travel.
  3. Registry (src/registry/) — named capability catalog (models, tools, agents, graphs, stores, middleware, policy) that .rag/.ragsh bind by name.
  4. Expressive language .rag (src/language/) — declarative, side-effect-free blueprint format that compiles into the same graph/harness runtime; the safe boundary for agent-authored plans.
  5. REPL language .ragsh (src/repl/) — imperative, capability-bound interactive orchestration; the RLM/CodeAct loop surface.

Page index

Getting started

  • Home — this page.
  • Quick Start — install, test, run an example, crate layout.
  • Examples — runnable graph, harness, provider, tool, structured output, sub-agent, and .rag examples.

Concepts

  • Recursion and RLM — the RLM execution model, the concrete recursive surfaces, and research lineage.
  • Architecture — how the five surfaces fit together.

Modules

Providers and contributing

  • Providers — OpenAI, Anthropic, Ollama, DeepSeek, Groq, xAI, OpenRouter, Together, Mistral, and OpenAI-compatible endpoints.
  • Development — building, testing, linting, and contributing.

Repository links

TinyAgents

Recursive language-model (RLM) harness for Rust.

Getting started

Concepts

Modules

Providers

Contributing

Clone this wiki locally