🤖🕸️ AgentMesh

Production-grade multi-agent orchestration framework powered by Pregel-style bulk-synchronous parallel (BSP) graph processing. Build sophisticated AI agent workflows with parallel execution, state management, and enterprise-grade observability.

Requires Go 1.24+

✨ Key Features

Core Engine

• 🔄 Pregel BSP Execution - Parallel graph processing with optimized concurrency (4-10x faster state access)
• 🧠 LLM Integration - Native support for OpenAI, Anthropic, Gemini, Amazon Bedrock, Ollama with streaming and reasoning models
• 💾 State Management - Lock-free channel-based state with checkpointing, managed value descriptors, and resume-time rehydration hooks
• 🛠️ Tool Orchestration - Type-safe function calling with automatic schema generation
• 🔀 Model Routing - Intelligent model selection based on cost, capabilities, and availability

Production Ready

• ✅ Graph Validation - Comprehensive compile-time error checking (cycles, missing nodes, unreachable paths)
• 💾 Checkpointing - Persistent state with auto-resume, encryption, and signing
• ♻️ Zero-Copy Resume - Copy-on-write checkpoint restores reuse the saved map and only allocate when keys mutate (10k+ key checkpoints resume without GC spikes)
• ⏸️ Human-in-the-Loop - Approval workflows with conditional guards and audit trails
• 📊 Observability - Built-in OpenTelemetry metrics, non-blocking event bus fan-out, and distributed tracing
• 🔁 Resilience - Configurable retry policies, circuit breakers, and timeouts
• 🔒 Security - WASM sandboxing and integrity checks

AI/ML Features

• 🔢 Embeddings & Memory - Semantic search and long-term conversation storage
• 🔍 RAG Integration - Vector stores (pgvector, Qdrant, Pinecone, Weaviate), AWS Bedrock Knowledge Bases, Kendra, and automatic query rephrasing
• 🧠 Native Reasoning - First-class support for o1, o3, Gemini 2.0, Claude reasoning models
• 📝 Prompt Templates - Variable substitution with reusable patterns

Extensibility

• 🤝 A2A Protocol - Multi-agent collaboration with standardized communication
• 🔌 MCP Support - Dynamic tool discovery from Model Context Protocol servers
• 🌐 LangChainGo Tools - Import existing tool ecosystem
• ⚙️ Custom Backends - Pluggable MessageBus for distributed execution (Redis, Kafka)

🚀 Quick Start

Installation

go get github.com/hupe1980/agentmesh@latest

Hello World ReAct Agent

package main

import (
    "context"
    "fmt"
    "log"
    "os"
    "strings"

    "github.com/hupe1980/agentmesh/pkg/agent"
    "github.com/hupe1980/agentmesh/pkg/graph"
    "github.com/hupe1980/agentmesh/pkg/message"
    "github.com/hupe1980/agentmesh/pkg/model/openai"
    "github.com/hupe1980/agentmesh/pkg/tool"
)

// WeatherArgs defines the JSON schema for the weather tool.
type WeatherArgs struct {
    Location string `json:"location" jsonschema:"description=The city to get weather for"`
}

func main() {
    ctx := context.Background()

    // Validate API key
    if strings.TrimSpace(os.Getenv("OPENAI_API_KEY")) == "" {
        log.Fatal("OPENAI_API_KEY environment variable is required")
    }

    // Create OpenAI model (uses OPENAI_API_KEY env var)
    model := openai.NewModel()

    // Define a tool with typed arguments
    weatherTool, err := tool.NewFuncTool(
        "get_weather",
        "Get current weather for a location",
        func(ctx context.Context, args WeatherArgs) (string, error) {
            return fmt.Sprintf("Weather in %s: Sunny, 72°F", args.Location), nil
        },
    )
    if err != nil {
        log.Fatal(err)
    }

    // Create ReAct agent
    reactAgent, err := agent.NewReAct(model,
        agent.WithTools(weatherTool),
        agent.WithMaxIterations(5),
    )
    if err != nil {
        log.Fatal(err)
    }

    // Execute agent and get the final result
    messages := []message.Message{
        message.NewHumanMessage("What's the weather in San Francisco?"),
    }

    lastMsg, err := graph.Last(reactAgent.Run(ctx, messages))
    if err != nil {
        log.Fatal(err)
    }

    fmt.Println(lastMsg.String())
}

Output:

Thought: I need to check the weather in San Francisco
Action: get_weather("San Francisco")
Observation: Weather in San Francisco: Sunny, 72°F
The weather in San Francisco is currently sunny with a temperature of 72°F.

📚 Documentation

Getting Started

• 📘 Getting Started Guide - Complete tutorial with examples
• 🏗️ Architecture Overview - Understanding the Pregel BSP design
• 📖 API Reference - Complete godoc

Core Concepts

• 🕸️ Graph Building - Nodes, edges, and execution flow
• 🗂️ State Management - Channels, reducers, checkpointing, and approval workflows
• 🔧 Tools Guide - Building and integrating tools
• 🤖 Model Integration - LLM provider setup and configuration

Advanced Features

• 🤖 Agent Patterns - ReAct, RAG, and Supervisor agents
• 📊 Observability - Metrics, tracing, and monitoring
• 🔌 Middleware - Caching, rate limiting, circuit breakers
• 📅 Custom Schedulers - Priority-based and resource-aware vertex execution
• 🧠 Memory & Embeddings - Semantic search and conversation storage
• 🤝 A2A Protocol - Multi-agent collaboration
• 🔒 WASM Sandboxing - Secure untrusted code execution

🎨 Examples

Explore 41 comprehensive examples in the examples/ directory:

See all examples →

Running Examples

# Run any example
cd examples/basic_agent
go run main.go

# Set required environment variables
export OPENAI_API_KEY=your-key-here
export ANTHROPIC_API_KEY=your-key-here  # For Anthropic examples

🏗️ Architecture

AgentMesh uses a layered architecture with clean separation of concerns:

┌─────────────────────────────────────────────────────────────┐
│           Application Layer (pkg/agent)                     │
│  • ReActAgent, SupervisorAgent, RAGAgent                    │
└────────────────────────┬────────────────────────────────────┘
                         │
                         ▼
┌─────────────────────────────────────────────────────────────┐
│           Graph Orchestration (pkg/graph)                   │
│  • Workflow construction • State management • Validation    │
└────────────────────────┬────────────────────────────────────┘
                         │
                         ▼
┌─────────────────────────────────────────────────────────────┐
│           Execution Engine (pkg/pregel)                     │
│                                                             │
│  • BSP Runtime      • Worker pools      • MessageBus        │
│  • Superstep sync   • Sharded frontier  • Backpressure      │
└─────────────────────────────────────────────────────────────┘

Key Components:

• Graph - Define nodes, edges, and execution flow with NodeFunc
• BSPState - Copy-on-write state with typed Key[T] and ListKey[T]
• Pregel Runtime - Bulk-synchronous parallel execution with sharded message passing
• Checkpointer - State persistence with encryption, signing, and two-phase commit
• Agents - High-level abstractions (ReAct, Supervisor, RAG)

Learn more about the architecture →

🧪 Testing

# Run all tests
go test ./...

# With coverage
go test ./... -coverprofile=coverage.out
go tool cover -html=coverage.out

# Run benchmarks
go test ./... -bench=. -benchmem

Testing Utilities

The pkg/testutil package provides mock implementations and builders for testing agents and workflows. See the Testing Guide for comprehensive documentation.

import "github.com/hupe1980/agentmesh/pkg/testutil"

// Create a mock model with builder pattern
model := testutil.NewModelBuilder().
    WithResponse("Hello!").
    WithToolCalls(message.ToolCall{Name: "search"}).
    Build()

// Create a mock tool
tool := testutil.NewToolBuilder("search").
    WithResult("search results").
    Build()

// Record and assert model interactions
recorder := testutil.NewConversationRecorder()
// ... run agent ...
recorder.AssertRequestCount(t, 1)
recorder.AssertToolCallMade(t, "search")

🤝 Contributing

Contributions are welcome!

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Write tests for your changes
4. Commit your changes (git commit -m 'Add amazing feature')
5. Push to the branch (git push origin feature/amazing-feature)
6. Open a Pull Request

Development Guidelines

• All new code must have tests (target 85%+ coverage)
• Run go fmt and golangci-lint before committing
• Update documentation for public API changes
• Add examples for new features

📄 License

Licensed under the Apache License 2.0 - see LICENSE for details.

🙏 Acknowledgments

• Pregel Paper - Google's bulk-synchronous parallel graph processing model
• Go Community - Exceptional tooling and ecosystem
• OpenTelemetry - Production-grade observability standards