Langgraphgo AI Agent

LangGraphGo

Abbreviated as lango, 中文: 懒狗

English | 简体中文

Website: http://lango.rpcx.io

🔀 Forked from paulnegz/langgraphgo - Enhanced with streaming, visualization, observability, and production-ready features.

This fork aims for feature parity with the Python LangGraph library, adding support for parallel execution, persistence, advanced state management, pre-built agents, and human-in-the-loop workflows.

🌐 Websites Built with LangGraphGo

Real-world applications built with LangGraphGo:

Insight	NoteX

Insight - An AI-powered knowledge management and insight generation platform that uses LangGraphGo to build intelligent analysis workflows, helping users extract key insights from massive amounts of information.

NoteX - An intelligent note-taking and knowledge organization tool that leverages AI for automatic categorization, tag extraction, and content association, making knowledge management more efficient.

📦 Installation

go get github.com/smallnest/langgraphgo

🚀 Features

• Core Runtime:

  • Parallel Execution: Concurrent node execution (fan-out) with thread-safe state merging.
  • Runtime Configuration: Propagate callbacks, tags, and metadata via RunnableConfig.
  • Generic Types: Type-safe state management with generic StateGraph implementations.
  • Native LLM Providers: Built-in nativeopenai and openairesponses providers with no langchaingo dependency, plus doubao and qwen.
  • LangChain Compatible: Framework-owned llms and tool types with a thin langchaingo adapter (llms.Wrap) for seamless interop.

• Persistence & Reliability:

  • Checkpointers: Redis, Postgres, SQLite, and File implementations for durable state.
  • File Checkpointing: Lightweight file-based checkpointing without external dependencies.
  • State Recovery: Pause and resume execution from checkpoints.

• Advanced Capabilities:

  • State Schema: Granular state updates with custom reducers (e.g., AppendReducer).
  • Smart Messages: Intelligent message merging with ID-based upserts (AddMessages).
  • Command API: Dynamic control flow and state updates directly from nodes.
  • Ephemeral Channels: Temporary state values that clear automatically after each step.
  • Subgraphs: Compose complex agents by nesting graphs within graphs.
  • Enhanced Streaming: Real-time event streaming with multiple modes (updates, values, messages).
  • Pre-built Agents: Ready-to-use ReAct, CreateAgent, and Supervisor agent factories.
  • Programmatic Tool Calling (PTC): LLM generates code that calls tools programmatically, reducing latency and token usage by 10x.

• Developer Experience:

  • Visualization: Export graphs to Mermaid, DOT, and ASCII with conditional edge support.
  • Human-in-the-loop (HITL): Interrupt execution, inspect state, edit history (UpdateState), and resume.
  • Observability: Built-in tracing and metrics support.
  • Tools: Integrated Tavily and Exa search tools.

🎯 Quick Start

package main

import (
	"context"
	"fmt"
	"log"

	"github.com/smallnest/langgraphgo/graph"
	openai "github.com/smallnest/langgraphgo/llms/nativeopenai"
)

func main() {
	ctx := context.Background()
	model, _ := openai.New()

	// 1. Create Graph
	g := graph.NewStateGraph[map[string]any]()

	// 2. Add Nodes
	g.AddNode("generate", "generate", func(ctx context.Context, state map[string]any) (map[string]any, error) {
		input, ok := state["input"].(string)
		if !ok {
			return nil, fmt.Errorf("invalid input")
		}

		response, err := model.Call(ctx, input)
		if err != nil {
			return nil, err
		}

		state["output"] = response
		return state, nil
	})

	// 3. Define Edges
	g.AddEdge("generate", graph.END)
	g.SetEntryPoint("generate")

	// 4. Compile
	runnable, _ := g.Compile()

	// 5. Invoke
	initialState := map[string]any{
		"input": "Hello, LangGraphGo!",
	}
	result, _ := runnable.Invoke(ctx, initialState)

	fmt.Println(result)
}

📚 Examples

This project includes 90+ comprehensive examples organized into categories:

Featured Examples

• ReAct Agent - Reason and Action agent using tools
• RAG Pipeline - Complete retrieval-augmented generation
• Chat Agent - Multi-turn conversation with session management
• Supervisor - Multi-agent orchestration
• Tree of Thoughts - Search-based reasoning with multiple solution paths
• Planning Agent - Dynamic workflow plan creation
• PEV Agent - Plan-Execute-Verify with self-correction
• Reflection Agent - Iterative improvement through self-reflection
• Mental Loop - Simulator-in-the-loop for safe action testing
• Reflexive Metacognitive Agent - Self-aware agent with explicit capabilities model

Example Categories

• Basic Concepts - Simple LLM integration, LangChain compatibility
• State Management - State schema, custom reducers, smart messages
• Graph Structure - Conditional routing, subgraphs, generics
• Parallel Execution - Fan-out/fan-in with state merging
• Streaming & Events - Real-time updates, listeners, logging
• Persistence - Checkpointing with file, memory, databases
• Human-in-the-Loop - Interrupts, approval, time travel
• Pre-built Agents - ReAct, Supervisor, Chat, Planning agents
• Programmatic Tool Calling - PTC for 10x latency reduction
• Memory - Buffer, sliding window, summarization strategies
• RAG - Vector stores, GraphRAG with FalkorDB
• Tools & Integrations - Search tools, GoSkills, MCP

View All 90+ Examples →

🔧 Key Concepts

Parallel Execution

LangGraphGo automatically executes nodes in parallel when they share the same starting node. Results are merged using the graph's state merger or schema.

g.AddEdge("start", "branch_a")
g.AddEdge("start", "branch_b")
// branch_a and branch_b run concurrently

Human-in-the-loop (HITL)

Pause execution to allow for human approval or input.

config := &graph.Config{
    InterruptBefore: []string{"human_review"},
}

// Execution stops before "human_review" node
state, err := runnable.InvokeWithConfig(ctx, input, config)

// Resume execution
resumeConfig := &graph.Config{
    ResumeFrom: []string{"human_review"},
}
runnable.InvokeWithConfig(ctx, state, resumeConfig)

Pre-built Agents

Quickly create complex agents using factory functions.

// Create a ReAct agent
agent, err := prebuilt.CreateReactAgent(model, tools)

// Create an agent with options
agent, err := prebuilt.CreateAgent(model, tools, prebuilt.WithSystemMessage("System prompt"))

// Create a Supervisor agent
supervisor, err := prebuilt.CreateSupervisor(model, agents)

Programmatic Tool Calling (PTC)

Generate code that calls tools directly, reducing API round-trips and token usage.

// Create a PTC agent
agent, err := ptc.CreatePTCAgent(ptc.PTCAgentConfig{
    Model:         model,
    Tools:         toolList,
    Language:      ptc.LanguagePython, // or ptc.LanguageGo
    ExecutionMode: ptc.ModeDirect,     // Subprocess (default) or ModeServer
    MaxIterations: 10,
})

// LLM generates code that calls tools programmatically
result, err := agent.Invoke(ctx, initialState)

See the PTC README for detailed documentation.

🎨 Graph Visualization

exporter := runnable.GetGraph()
fmt.Println(exporter.DrawMermaid()) // Generates Mermaid flowchart

📈 Performance

• Graph Operations: ~14-94μs depending on format
• Tracing Overhead: ~4μs per execution
• Event Processing: 1000+ events/second
• Streaming Latency: <100ms

🧪 Testing

go test ./... -v

Contributors

This project is open for contributions! if you are interested in being a contributor please create feature issues first, then submit PRs..

📄 License

MIT License - see original repository for details.