AutoAgents

A production-grade multi-agent framework in Rust

English | 中文 | 日本語 | Español | Français | Deutsch | 한국어 | Português (Brasil)
_{Translations may lag behind the English README.}

Documentation | Examples | Contributing

Like this project? Star us on GitHub

AutoAgents is a modular, multi-agent framework for building intelligent systems in Rust. It combines a type-safe agent model with structured tool calling, configurable memory, and pluggable LLM backends. The architecture is designed for performance, safety, and composability across server and edge, and serves as the foundation for higher-level systems like Odyssey.

• Agent execution: ReAct and basic executors, streaming responses, and structured outputs
• Tooling: Derive macros for tools and outputs, plus a sandboxed WASM runtime for tool execution
• Memory: Sliding window memory with extensible backends
• LLM providers: Cloud and local backends behind a unified interface
• LLM Guardrails: Guardrail implementation for safeguarding LLM inference
• LLM Optimization: Build LLM pipelines with optimization passes like cache and retry for faster, more reliable inference
• Multi-agent orchestration: Typed pub/sub communication and environment management
• Speech-Processing: Local TTS and STT support
• Observability: OpenTelemetry tracing and metrics with pluggable exporters

See https://github.com/liquidos-ai/AutoAgents-Experimental-Backends

Provider support is actively expanding based on community needs.

• Rust (latest stable recommended)
• Cargo package manager
• LeftHook for Git hooks management
• Python 3.9+ (required for Python bindings)
• uv for Python environment and package management
• maturin (required to build/install local Python bindings from source)

sudo apt update
sudo apt install build-essential libasound2-dev alsa-utils pkg-config libssl-dev -y

macOS (Homebrew):

brew install lefthook

Linux/Windows (npm):

npm install -g lefthook

git clone https://github.com/liquidos-ai/AutoAgents.git
cd AutoAgents
lefthook install
cargo build --workspace --all-features

AutoAgents ships Python bindings to PyPI. Install the base package and add backends via extras:

pip install autoagents-py                            # core + cloud LLM providers
pip install "autoagents-py[llamacpp]"               # + llama.cpp CPU
pip install "autoagents-py[llamacpp-cuda]"          # + llama.cpp CUDA
pip install "autoagents-py[llamacpp-metal]"         # + llama.cpp Metal (macOS)
pip install "autoagents-py[llamacpp-vulkan]"        # + llama.cpp Vulkan
pip install "autoagents-py[mistralrs]"              # + mistral-rs CPU
pip install "autoagents-py[mistralrs-cuda]"         # + mistral-rs CUDA
pip install "autoagents-py[mistralrs-metal]"        # + mistral-rs Metal (macOS)
pip install "autoagents-py[guardrails]"             # + Guardrails
pip install "autoagents-py[llamacpp-cuda,guardrails]"  # combine extras

Development install from this repo:

uv venv --python=3.12
source .venv/bin/activate          # Windows: .venv\Scripts\activate
uv pip install -U pip maturin pytest pytest-asyncio pytest-cov

# Clean, build, and install all CPU bindings into the active venv
make python-bindings-build

# Clean, build, and install CPU + CUDA bindings
make python-bindings-build-cuda

The Make targets remove stale editable-install extension artifacts before rebuilding, which avoids loading out-of-date .abi3.so files from the source tree.

Example scripts:

• Core cloud example: bindings/python/autoagents/examples/openai_agent.py
• llama.cpp example: bindings/python/autoagents-llamacpp/examples/llamacpp_agent.py
• mistral-rs example: bindings/python/autoagents-mistralrs/examples/mistral_rs_agent.py

cargo test --features "full" --workspace

use autoagents::core::agent::memory::SlidingWindowMemory;
use autoagents::core::agent::prebuilt::executor::{ReActAgent, ReActAgentOutput};
use autoagents::core::agent::task::Task;
use autoagents::core::agent::{AgentBuilder, AgentDeriveT, AgentOutputT, DirectAgent};
use autoagents::core::error::Error;
use autoagents::core::tool::{ToolCallError, ToolInputT, ToolRuntime, ToolT};
use autoagents::llm::LLMProvider;
use autoagents::llm::backends::openai::OpenAI;
use autoagents::llm::builder::LLMBuilder;
use autoagents_derive::{agent, tool, AgentHooks, AgentOutput, ToolInput};
use serde::{Deserialize, Serialize};
use serde_json::Value;
use std::sync::Arc;

#[derive(Serialize, Deserialize, ToolInput, Debug)]
pub struct AdditionArgs {
    #[input(description = "Left Operand for addition")]
    left: i64,
    #[input(description = "Right Operand for addition")]
    right: i64,
}

#[tool(
    name = "Addition",
    description = "Use this tool to Add two numbers",
    input = AdditionArgs,
)]
struct Addition {}

#[async_trait]
impl ToolRuntime for Addition {
    async fn execute(&self, args: Value) -> Result<Value, ToolCallError> {
        println!("execute tool: {:?}", args);
        let typed_args: AdditionArgs = serde_json::from_value(args)?;
        let result = typed_args.left + typed_args.right;
        Ok(result.into())
    }
}

#[derive(Debug, Serialize, Deserialize, AgentOutput)]
pub struct MathAgentOutput {
    #[output(description = "The addition result")]
    value: i64,
    #[output(description = "Explanation of the logic")]
    explanation: String,
    #[output(description = "If user asks other than math questions, use this to answer them.")]
    generic: Option<String>,
}

#[agent(
    name = "math_agent",
    description = "You are a Math agent",
    tools = [Addition],
    output = MathAgentOutput,
)]
#[derive(Default, Clone, AgentHooks)]
pub struct MathAgent {}

impl From<ReActAgentOutput> for MathAgentOutput {
    fn from(output: ReActAgentOutput) -> Self {
        let resp = output.response;
        if output.done && !resp.trim().is_empty() {
            if let Ok(value) = serde_json::from_str::<MathAgentOutput>(&resp) {
                return value;
            }
        }
        MathAgentOutput {
            value: 0,
            explanation: resp,
            generic: None,
        }
    }
}

pub async fn simple_agent(llm: Arc<dyn LLMProvider>) -> Result<(), Error> {
    let sliding_window_memory = Box::new(SlidingWindowMemory::new(10));

    let agent_handle = AgentBuilder::<_, DirectAgent>::new(ReActAgent::new(MathAgent {}))
        .llm(llm)
        .memory(sliding_window_memory)
        .build()
        .await?;

    let result = agent_handle.agent.run(Task::new("What is 1 + 1?")).await?;
    println!("Result: {:?}", result);
    Ok(())
}

#[tokio::main]
async fn main() -> Result<(), Error> {
    let api_key = std::env::var("OPENAI_API_KEY").unwrap_or("".into());

    let llm: Arc<OpenAI> = LLMBuilder::<OpenAI>::new()
        .api_key(api_key)
        .model("gpt-4o")
        .max_tokens(512)
        .temperature(0.2)
        .build()
        .expect("Failed to build LLM");

    let _ = simple_agent(llm).await?;
    Ok(())
}

AutoAgents CLI helps in running Agentic Workflows from YAML configurations and serves them over HTTP. You can check it out at https://github.com/liquidos-ai/AutoAgents-CLI.

Explore the examples to get started quickly:

Demonstrates various examples like Simple Agent with Tools, Very Basic Agent, Edge Agent, Chaining, Actor Based Model, Streaming and Adding Agent Hooks.

Demonstrates LLM pipelines with optimization passes such as cache and retry to improve performance and reliability.

Demonstrates configurable input and output guardrails with Block, Sanitize, and Audit policies using an LLMLayer in the pipeline.

Demonstrates how to integrate AutoAgents with the Model Context Protocol (MCP).

Demonstrates how to integrate AutoAgents with the Mistral-rs for Local Models.

Demonstrates various design patterns like Chaining, Planning, Routing, Parallel and Reflection.

Contains examples demonstrating how to use different LLM providers with AutoAgents.

A simple agent which can run tools in WASM runtime.

A sophisticated ReAct-based coding agent with file manipulation capabilities.

Run AutoAgents Speech Example with realtime TTS and STT.

Example App that runs AutoAgents with Local models in Android using AutoAgents-llamacpp backend

AutoAgents is built with a modular architecture:

AutoAgents/
├── crates/
│   ├── autoagents/                # Main library entry point
│   ├── autoagents-core/           # Core agent framework
│   ├── autoagents-protocol/       # Shared protocol/event types
│   ├── autoagents-llm/            # LLM provider implementations
│   ├── autoagents-telemetry/      # OpenTelemetry integration
│   ├── autoagents-toolkit/        # Collection of ready-to-use tools
│   ├── autoagents-mistral-rs/     # LLM provider implementations using Mistral-rs
│   ├── autoagents-llamacpp/       # LLM provider implementation using LlamaCpp
│   ├── autoagents-speech/         # Speech model support for TTS and STT
│   ├── autoagents-guardrails/     # LLM Guardrails implementation
│   ├── autoagents-qdrant/         # Qdrant vector store
│   └── autoagents-derive/         # Procedural macros
├── examples/                      # Example implementations
├── bindings/                      # Bindings for different languages

• Agent: The fundamental unit of intelligence
• Environment: Manages agent lifecycle and communication
• Memory: Configurable memory systems
• Tools: External capability integration
• Executors: Different reasoning patterns (ReAct, Chain-of-Thought)

sudo apt update
sudo apt install build-essential libasound2-dev alsa-utils pkg-config libssl-dev -y

cargo test --workspace --features default --exclude autoagents-burn --exclude autoagents-mistral-rs --exclude wasm_agent

# Coverage (requires cargo-tarpaulin)
cargo install cargo-tarpaulin
cargo tarpaulin --all-features --out html

cargo bench -p autoagents-core --bench agent_runtime

This project uses LeftHook for Git hooks management. The hooks will automatically:

• Format code with cargo fmt --check
• Run linting with cargo clippy -- -D warnings
• Execute tests with cargo test --all-features --workspace --exclude autoagents-burn

We welcome contributions. Please see our Contributing Guidelines and Code of Conduct for details.

• API Documentation: Complete framework docs
• Examples: Practical implementation examples

• GitHub Issues: Bug reports and feature requests
• Discussions: Community Q&A and ideas
• Discord: Join our Discord Community using https://discord.gg/zfAF9MkEtK

AutoAgents is designed for high performance:

• Memory Efficient: Optimized memory usage with configurable backends
• Concurrent: Full async/await support with tokio
• Scalable: Horizontal scaling with multi-agent coordination
• Type Safe: Compile-time guarantees with Rust's type system

AutoAgents is dual-licensed under:

• MIT License (MIT_LICENSE)
• Apache License 2.0 (APACHE_LICENSE)

You may choose either license for your use case.

Built by the Liquidos AI team and wonderful community of researchers and engineers.

Special thanks to:

• The Rust community for the excellent ecosystem
• LLM providers for enabling high-quality model APIs
• All contributors who help improve AutoAgents