Graph Agents

Build complex, stateful workflows using LangGraph-style orchestration with native ADK-Rust integration.

Overview

GraphAgent allows you to define workflows as directed graphs with nodes and edges, supporting:

• AgentNode: Wrap LLM agents as graph nodes with custom input/output mappers
• Cyclic Workflows: Native support for loops and iterative reasoning (ReAct pattern)
• Conditional Routing: Dynamic edge routing based on state
• State Management: Typed state with reducers (overwrite, append, sum, custom)
• Checkpointing: Persistent state for fault tolerance and human-in-the-loop
• Streaming: Multiple stream modes (values, updates, messages, debug)

The adk-graph crate provides LangGraph-style workflow orchestration for building complex, stateful agent workflows. It brings graph-based workflow capabilities to the ADK-Rust ecosystem while maintaining full compatibility with ADK's agent system.

Key Benefits:

• Visual Workflow Design: Define complex logic as intuitive node-and-edge graphs
• Parallel Execution: Multiple nodes can run simultaneously for better performance
• State Persistence: Built-in checkpointing for fault tolerance and human-in-the-loop
• LLM Integration: Native support for wrapping ADK agents as graph nodes
• Flexible Routing: Static edges, conditional routing, and dynamic decision making

Choosing Between the Workflow Agents and the Graph

ADK-Rust supports two ways to orchestrate. Neither replaces the other, and both are maintained.

Prefer the workflow agents when the shape of the work is known and linear: they are shorter to write and there is no state schema to maintain. Reach for the graph when control flow depends on results, or when a run has to outlive the process.

They compose

The three workflow agents implement Agent, and AgentNode wraps any Agent, so a workflow agent is a graph node:

use adk_agent::SequentialAgent;
use adk_graph::node::AgentNode;
use std::sync::Arc;

let pipeline = Arc::new(SequentialAgent::new("pipeline", vec![extract, validate]));
let node = AgentNode::new(pipeline as Arc<dyn adk_core::Agent>);
// `node` now goes into a StateGraph like any other node.

GraphAgent also implements Agent, so the reverse holds: a graph can be a sub-agent of a SequentialAgent. Use the graph for the part that needs branching or durability, and the workflow agents for the parts that do not.

What You'll Build

In this guide, you'll create a Text Processing Pipeline that runs translation and summarization in parallel:

                        ┌─────────────────────┐
       User Input       │                     │
      ────────────────▶ │       START         │
                        │                     │
                        └──────────┬──────────┘
                                   │
                   ┌───────────────┴───────────────┐
                   │                               │
                   ▼                               ▼
        ┌──────────────────┐            ┌──────────────────┐
        │   TRANSLATOR     │            │   SUMMARIZER     │
        │                  │            │                  │
        │  🇫🇷 French       │            │  📝 One sentence │
        │     Translation  │            │     Summary      │
        └─────────┬────────┘            └─────────┬────────┘
                  │                               │
                  └───────────────┬───────────────┘
                                  │
                                  ▼
                        ┌─────────────────────┐
                        │      COMBINE        │
                        │                     │
                        │  📋 Merge Results   │
                        └──────────┬──────────┘
                                   │
                                   ▼
                        ┌─────────────────────┐
                        │        END          │
                        │                     │
                        │   ✅ Complete       │
                        └─────────────────────┘

Key Concepts:

• Nodes - Processing units that perform work (LLM agents, functions, or custom logic)
• Edges - Control flow between nodes (static connections or conditional routing)
• State - Shared data that flows through the graph and persists between nodes
• Parallel Execution - Multiple nodes can run simultaneously for better performance

Understanding the Core Components

🔧 Nodes: The Workers Nodes are where the actual work happens. Each node can:

• AgentNode: Wrap an LLM agent to process natural language
• Function Node: Execute custom Rust code for data processing
• Built-in Nodes: Use predefined logic like counters or validators

Think of nodes as specialized workers in an assembly line - each has a specific job and expertise.

🔀 Edges: The Flow Control Edges determine how execution moves through your graph:

• Static Edges: Direct connections (A → B → C)
• Conditional Edges: Dynamic routing based on state (if sentiment == "positive" → positive_handler)
• Parallel Edges: Multiple paths from one node (START → [translator, summarizer])

Edges are like traffic signals and road signs that direct the flow of work.

💾 State: The Shared Memory State is a key-value store that all nodes can read from and write to:

• Input Data: Initial information fed into the graph
• Intermediate Results: Output from one node becomes input for another
• Final Output: The completed result after all processing

State acts like a shared whiteboard where nodes can leave information for others to use.

⚡ Parallel Execution: The Speed Boost When multiple edges leave a node, those target nodes run simultaneously:

• Faster Processing: Independent tasks run at the same time
• Resource Efficiency: Better utilization of CPU and I/O
• Scalability: Handle more complex workflows without linear slowdown

This is like having multiple workers tackle different parts of a job simultaneously instead of waiting in line.

---

Quick Start

1. Create Your Project

cargo new graph_demo
cd graph_demo

Add dependencies to Cargo.toml:

[dependencies]
adk-graph = { version = "2.1.0", features = ["sqlite"] }
adk-agent = "2.1.0"
adk-model = "2.1.0"
adk-core = "2.1.0"
tokio = { version = "1", features = ["full"] }
dotenvy = "0.15"
serde_json = "1.0"

Create .env with your API key:

echo 'GOOGLE_API_KEY=your-api-key' > .env

2. Parallel Processing Example

Here's a complete working example that processes text in parallel:

use adk_agent::LlmAgentBuilder;
use adk_graph::{
    agent::GraphAgent,
    edge::{END, START},
    node::{AgentNode, ExecutionConfig, NodeOutput},
    state::State,
};
use adk_model::GeminiModel;
use serde_json::json;
use std::sync::Arc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    dotenvy::dotenv().ok();
    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = Arc::new(GeminiModel::new(&api_key, "gemini-3.7-flash")?);

    // Create specialized LLM agents
    let translator_agent = Arc::new(
        LlmAgentBuilder::new("translator")
            .description("Translates text to French")
            .model(model.clone())
            .instruction("Translate the input text to French. Only output the translation.")
            .build()?,
    );

    let summarizer_agent = Arc::new(
        LlmAgentBuilder::new("summarizer")
            .description("Summarizes text")
            .model(model.clone())
            .instruction("Summarize the input text in one sentence.")
            .build()?,
    );

    // Wrap agents as graph nodes with input/output mappers
    let translator_node = AgentNode::new(translator_agent)
        .with_input_mapper(|state| {
            let text = state.get("input").and_then(|v| v.as_str()).unwrap_or("");
            adk_core::Content::new("user").with_text(text)
        })
        .with_output_mapper(|events| {
            let mut updates = std::collections::HashMap::new();
            for event in events {
                if let Some(content) = event.content() {
                    let text: String = content.parts.iter()
                        .filter_map(|p| p.text())
                        .collect::<Vec<_>>()
                        .join("");
                    if !text.is_empty() {
                        updates.insert("translation".to_string(), json!(text));
                    }
                }
            }
            updates
        });

    let summarizer_node = AgentNode::new(summarizer_agent)
        .with_input_mapper(|state| {
            let text = state.get("input").and_then(|v| v.as_str()).unwrap_or("");
            adk_core::Content::new("user").with_text(text)
        })
        .with_output_mapper(|events| {
            let mut updates = std::collections::HashMap::new();
            for event in events {
                if let Some(content) = event.content() {
                    let text: String = content.parts.iter()
                        .filter_map(|p| p.text())
                        .collect::<Vec<_>>()
                        .join("");
                    if !text.is_empty() {
                        updates.insert("summary".to_string(), json!(text));
                    }
                }
            }
            updates
        });

    // Build the graph with parallel execution
    let agent = GraphAgent::builder("text_processor")
        .description("Processes text with translation and summarization in parallel")
        .channels(&["input", "translation", "summary", "result"])
        .node(translator_node)
        .node(summarizer_node)
        .node_fn("combine", |ctx| async move {
            let translation = ctx.get("translation").and_then(|v| v.as_str()).unwrap_or("N/A");
            let summary = ctx.get("summary").and_then(|v| v.as_str()).unwrap_or("N/A");

            let result = format!(
                "=== Processing Complete ===\n\n\
                French Translation:\n{}\n\n\
                Summary:\n{}",
                translation, summary
            );

            Ok(NodeOutput::new().with_update("result", json!(result)))
        })
        // Parallel execution: both nodes start simultaneously
        .edge(START, "translator")
        .edge(START, "summarizer")
        .edge("translator", "combine")
        .edge("summarizer", "combine")
        .edge("combine", END)
        .build()?;

    // Execute the graph
    let mut input = State::new();
    input.insert("input".to_string(), json!("AI is transforming how we work and live."));

    let result = agent.invoke(input, ExecutionConfig::new("thread-1")).await?;
    println!("{}", result.get("result").and_then(|v| v.as_str()).unwrap_or(""));

    Ok(())
}

Example Output:

=== Processing Complete ===

French Translation:
L'IA transforme notre façon de travailler et de vivre.

Summary:
AI is revolutionizing work and daily life through technological transformation.

How Graph Execution Works

The Big Picture

Graph agents execute in super-steps - all ready nodes run in parallel, then the graph waits for all to complete before the next step:

Step 1: START ──┬──▶ translator (running)
                └──▶ summarizer (running)
                
                ⏳ Wait for both to complete...
                
Step 2: translator ──┬──▶ combine (running)
        summarizer ──┘
        
                ⏳ Wait for combine to complete...
                
Step 3: combine ──▶ END ✅

State Flow Through Nodes

Each node can read from and write to the shared state:

┌─────────────────────────────────────────────────────────────────────┐
│ STEP 1: Initial state                                               │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│   State: { "input": "AI is transforming how we work" }             │
│                                                                     │
│                              ↓                                      │
│                                                                     │
│   ┌──────────────────┐              ┌──────────────────┐           │
│   │   translator     │              │   summarizer     │           │
│   │  reads "input"   │              │  reads "input"   │           │
│   └──────────────────┘              └──────────────────┘           │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────────┐
│ STEP 2: After parallel execution                                    │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│   State: {                                                          │
│     "input": "AI is transforming how we work",                     │
│     "translation": "L'IA transforme notre façon de travailler",    │
│     "summary": "AI is revolutionizing work through technology"     │
│   }                                                                 │
│                                                                     │
│                              ↓                                      │
│                                                                     │
│   ┌──────────────────────────────────────┐                         │
│   │           combine                    │                         │
│   │  reads "translation" + "summary"     │                         │
│   │  writes "result"                     │                         │
│   └──────────────────────────────────────┘                         │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────────┐
│ STEP 3: Final state                                                 │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│   State: {                                                          │
│     "input": "AI is transforming how we work",                     │
│     "translation": "L'IA transforme notre façon de travailler",    │
│     "summary": "AI is revolutionizing work through technology",    │
│     "result": "=== Processing Complete ===\n\nFrench..."          │
│   }                                                                 │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

What Makes It Work

---

Conditional Routing with LLM Classification

Build smart routing systems where LLMs decide the execution path:

Visual: Sentiment-Based Routing

                        ┌─────────────────────┐
       User Feedback    │                     │
      ────────────────▶ │    CLASSIFIER       │
                        │  🧠 Analyze tone    │
                        └──────────┬──────────┘
                                   │
                   ┌───────────────┼───────────────┐
                   │               │               │
                   ▼               ▼               ▼
        ┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐
        │   POSITIVE       │ │    NEGATIVE      │ │    NEUTRAL       │
        │                  │ │                  │ │                  │
        │  😊 Thank you!   │ │  😔 Apologize    │ │  😐 Ask more     │
        │     Celebrate    │ │     Help fix     │ │     questions    │
        └──────────────────┘ └──────────────────┘ └──────────────────┘

Complete Example Code

use adk_agent::LlmAgentBuilder;
use adk_graph::{
    edge::{END, Router, START},
    graph::StateGraph,
    node::{AgentNode, ExecutionConfig},
    state::State,
};
use adk_model::GeminiModel;
use serde_json::json;
use std::sync::Arc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    dotenvy::dotenv().ok();
    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = Arc::new(GeminiModel::new(&api_key, "gemini-3.7-flash")?);

    // Create classifier agent
    let classifier_agent = Arc::new(
        LlmAgentBuilder::new("classifier")
            .description("Classifies text sentiment")
            .model(model.clone())
            .instruction(
                "You are a sentiment classifier. Analyze the input text and respond with \
                ONLY one word: 'positive', 'negative', or 'neutral'. Nothing else.",
            )
            .build()?,
    );

    // Create response agents for each sentiment
    let positive_agent = Arc::new(
        LlmAgentBuilder::new("positive")
            .description("Handles positive feedback")
            .model(model.clone())
            .instruction(
                "You are a customer success specialist. The customer has positive feedback. \
                Express gratitude, reinforce the positive experience, and suggest ways to \
                share their experience. Be warm and appreciative. Keep response under 3 sentences.",
            )
            .build()?,
    );

    let negative_agent = Arc::new(
        LlmAgentBuilder::new("negative")
            .description("Handles negative feedback")
            .model(model.clone())
            .instruction(
                "You are a customer support specialist. The customer has a complaint. \
                Acknowledge their frustration, apologize sincerely, and offer help. \
                Be empathetic. Keep response under 3 sentences.",
            )
            .build()?,
    );

    let neutral_agent = Arc::new(
        LlmAgentBuilder::new("neutral")
            .description("Handles neutral feedback")
            .model(model.clone())
            .instruction(
                "You are a customer service representative. The customer has neutral feedback. \
                Ask clarifying questions to better understand their needs. Be helpful and curious. \
                Keep response under 3 sentences.",
            )
            .build()?,
    );

    // Create AgentNodes with mappers
    let classifier_node = AgentNode::new(classifier_agent)
        .with_input_mapper(|state| {
            let text = state.get("feedback").and_then(|v| v.as_str()).unwrap_or("");
            adk_core::Content::new("user").with_text(text)
        })
        .with_output_mapper(|events| {
            let mut updates = std::collections::HashMap::new();
            for event in events {
                if let Some(content) = event.content() {
                    let text: String = content.parts.iter()
                        .filter_map(|p| p.text())
                        .collect::<Vec<_>>()
                        .join("")
                        .to_lowercase()
                        .trim()
                        .to_string();
                    
                    let sentiment = if text.contains("positive") { "positive" }
                        else if text.contains("negative") { "negative" }
                        else { "neutral" };
                    
                    updates.insert("sentiment".to_string(), json!(sentiment));
                }
            }
            updates
        });

    // Response nodes (similar pattern for each)
    let positive_node = AgentNode::new(positive_agent)
        .with_input_mapper(|state| {
            let text = state.get("feedback").and_then(|v| v.as_str()).unwrap_or("");
            adk_core::Content::new("user").with_text(text)
        })
        .with_output_mapper(|events| {
            let mut updates = std::collections::HashMap::new();
            for event in events {
                if let Some(content) = event.content() {
                    let text: String = content.parts.iter()
                        .filter_map(|p| p.text())
                        .collect::<Vec<_>>()
                        .join("");
                    updates.insert("response".to_string(), json!(text));
                }
            }
            updates
        });

    // Build graph with conditional routing
    let graph = StateGraph::with_channels(&["feedback", "sentiment", "response"])
        .add_node(classifier_node)
        .add_node(positive_node)
        // ... add negative_node and neutral_node similarly
        .add_edge(START, "classifier")
        .add_conditional_edges(
            "classifier",
            Router::by_field("sentiment"),  // Route based on sentiment field
            [
                ("positive", "positive"),
                ("negative", "negative"),
                ("neutral", "neutral"),
            ],
        )
        .add_edge("positive", END)
        .add_edge("negative", END)
        .add_edge("neutral", END)
        .compile()?;

    // Test with different feedback
    let mut input = State::new();
    input.insert("feedback".to_string(), json!("Your product is amazing! I love it!"));

    let result = graph.invoke(input, ExecutionConfig::new("feedback-1")).await?;
    println!("Sentiment: {}", result.get("sentiment").and_then(|v| v.as_str()).unwrap_or(""));
    println!("Response: {}", result.get("response").and_then(|v| v.as_str()).unwrap_or(""));

    Ok(())
}

Example Flow:

Input: "Your product is amazing! I love it!"
       ↓
Classifier: "positive"
       ↓
Positive Agent: "Thank you so much for the wonderful feedback! 
                We're thrilled you love our product. 
                Would you consider leaving a review to help others?"

ReAct Pattern: Reasoning + Acting

Build agents that can use tools iteratively to solve complex problems:

Visual: ReAct Cycle

                        ┌─────────────────────┐
       User Question    │                     │
      ────────────────▶ │      REASONER       │
                        │  🧠 Think + Act     │
                        └──────────┬──────────┘
                                   │
                                   ▼
                        ┌─────────────────────┐
                        │   Has tool calls?   │
                        │                     │
                        └──────────┬──────────┘
                                   │
                   ┌───────────────┴───────────────┐
                   │                               │
                   ▼                               ▼
        ┌──────────────────┐            ┌──────────────────┐
        │       YES        │            │        NO        │
        │                  │            │                  │
        │  🔄 Loop back    │            │  ✅ Final answer │
        │     to reasoner  │            │      END         │
        └─────────┬────────┘            └──────────────────┘
                  │
                  └─────────────────┐
                                    │
                                    ▼
                        ┌─────────────────────┐
                        │      REASONER       │
                        │  🧠 Think + Act     │
                        │   (next iteration)  │
                        └─────────────────────┘

Complete ReAct Example

use adk_agent::LlmAgentBuilder;
use adk_core::{Part, Tool};
use adk_graph::{
    edge::{END, START},
    graph::StateGraph,
    node::{AgentNode, ExecutionConfig, NodeOutput},
    state::State,
};
use adk_model::GeminiModel;
use adk_tool::FunctionTool;
use serde_json::json;
use std::sync::Arc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    dotenvy::dotenv().ok();
    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = Arc::new(GeminiModel::new(&api_key, "gemini-3.7-flash")?);

    // Create tools
    let weather_tool = Arc::new(FunctionTool::new(
        "get_weather",
        "Get the current weather for a location. Takes a 'location' parameter (city name).",
        |_ctx, args| async move {
            let location = args.get("location").and_then(|v| v.as_str()).unwrap_or("unknown");
            Ok(json!({
                "location": location,
                "temperature": "72°F",
                "condition": "Sunny",
                "humidity": "45%"
            }))
        },
    )) as Arc<dyn Tool>;

    let calculator_tool = Arc::new(FunctionTool::new(
        "calculator",
        "Perform mathematical calculations. Takes an 'expression' parameter (string).",
        |_ctx, args| async move {
            let expr = args.get("expression").and_then(|v| v.as_str()).unwrap_or("0");
            let result = match expr {
                "2 + 2" => "4",
                "10 * 5" => "50",
                "100 / 4" => "25",
                "15 - 7" => "8",
                _ => "Unable to evaluate",
            };
            Ok(json!({ "result": result, "expression": expr }))
        },
    )) as Arc<dyn Tool>;

    // Create reasoner agent with tools
    let reasoner_agent = Arc::new(
        LlmAgentBuilder::new("reasoner")
            .description("Reasoning agent with tools")
            .model(model.clone())
            .instruction(
                "You are a helpful assistant with access to tools. Use tools when needed to answer questions. \
                When you have enough information, provide a final answer without using more tools.",
            )
            .tool(weather_tool)
            .tool(calculator_tool)
            .build()?,
    );

    // Create reasoner node that detects tool usage
    let reasoner_node = AgentNode::new(reasoner_agent)
        .with_input_mapper(|state| {
            let question = state.get("question").and_then(|v| v.as_str()).unwrap_or("");
            adk_core::Content::new("user").with_text(question)
        })
        .with_output_mapper(|events| {
            let mut updates = std::collections::HashMap::new();
            let mut has_tool_calls = false;
            let mut response = String::new();

            for event in events {
                if let Some(content) = event.content() {
                    for part in &content.parts {
                        match part {
                            Part::FunctionCall { .. } => {
                                has_tool_calls = true;
                            }
                            Part::Text { text } => {
                                response.push_str(text);
                            }
                            _ => {}
                        }
                    }
                }
            }

            updates.insert("has_tool_calls".to_string(), json!(has_tool_calls));
            updates.insert("response".to_string(), json!(response));
            updates
        });

    // Build ReAct graph with cycle
    let graph = StateGraph::with_channels(&["question", "has_tool_calls", "response", "iteration"])
        .add_node(reasoner_node)
        .add_node_fn("counter", |ctx| async move {
            let i = ctx.get("iteration").and_then(|v| v.as_i64()).unwrap_or(0);
            Ok(NodeOutput::new().with_update("iteration", json!(i + 1)))
        })
        .add_edge(START, "counter")
        .add_edge("counter", "reasoner")
        .add_conditional_edges(
            "reasoner",
            |state| {
                let has_tools = state.get("has_tool_calls").and_then(|v| v.as_bool()).unwrap_or(false);
                let iteration = state.get("iteration").and_then(|v| v.as_i64()).unwrap_or(0);

                // Safety limit
                if iteration >= 5 { return END.to_string(); }

                if has_tools {
                    "counter".to_string()  // Loop back for more reasoning
                } else {
                    END.to_string()  // Done - final answer
                }
            },
            [("counter", "counter"), (END, END)],
        )
        .compile()?
        .with_recursion_limit(10);

    // Test the ReAct agent
    let mut input = State::new();
    input.insert("question".to_string(), json!("What's the weather in Paris and what's 15 + 25?"));

    let result = graph.invoke(input, ExecutionConfig::new("react-1")).await?;
    println!("Final answer: {}", result.get("response").and_then(|v| v.as_str()).unwrap_or(""));
    println!("Iterations: {}", result.get("iteration").and_then(|v| v.as_i64()).unwrap_or(0));

    Ok(())
}

Example Flow:

Question: "What's the weather in Paris and what's 15 + 25?"

Iteration 1:
  Reasoner: "I need to get weather info and do math"
  → Calls get_weather(location="Paris") and calculator(expression="15 + 25")
  → has_tool_calls = true → Loop back

Iteration 2:
  Reasoner: "Based on the results: Paris is 72°F and sunny, 15 + 25 = 40"
  → No tool calls → has_tool_calls = false → END

Final Answer: "The weather in Paris is 72°F and sunny with 45% humidity. 
              And 15 + 25 equals 40."

AgentNode

Wraps any ADK Agent (typically LlmAgent) as a graph node:

What the agent sees

An agent inside a graph runs under a context derived from the invocation that started the graph, so it behaves the same as it does outside one. When a Runner invokes a GraphAgent, the caller's identity and services are carried through automatically:

A graph invoked directly — graph.invoke(state, ExecutionConfig::new("thread")) — has no invocation to inherit. That is standalone mode: the node gets user_id = "graph_user", app_name = "graph_app", branch main, no secrets, and no memory. It is a deliberate mode for running a graph outside a Runner, not a fallback to reach for in production.

To bridge manually — for instance when driving a graph from your own executor — pass the invocation explicitly:

let config = ExecutionConfig::new(ctx.session_id()).with_parent_context(ctx.clone());

Note: the node still runs on its own in-memory graph session, so agent conversation history inside a node is scoped to the node rather than appended to the caller's session.

let node = AgentNode::new(llm_agent)
    .with_input_mapper(|state| {
        // Transform graph state to agent input Content
        let text = state.get("input").and_then(|v| v.as_str()).unwrap_or("");
        adk_core::Content::new("user").with_text(text)
    })
    .with_output_mapper(|events| {
        // Transform agent events to state updates
        let mut updates = std::collections::HashMap::new();
        for event in events {
            if let Some(content) = event.content() {
                let text: String = content.parts.iter()
                    .filter_map(|p| p.text())
                    .collect::<Vec<_>>()
                    .join("");
                updates.insert("output".to_string(), json!(text));
            }
        }
        updates
    });

Function Nodes

Simple async functions that process state:

.node_fn("process", |ctx| async move {
    let input = ctx.state.get("input").unwrap();
    let output = process_data(input).await?;
    Ok(NodeOutput::new().with_update("output", output))
})

Edge Types

Static Edges

Direct connections between nodes:

.edge(START, "first_node")
.edge("first_node", "second_node")
.edge("second_node", END)

Conditional Edges

Dynamic routing based on state:

.conditional_edge(
    "router",
    |state| {
        match state.get("next").and_then(|v| v.as_str()) {
            Some("research") => "research_node".to_string(),
            Some("write") => "write_node".to_string(),
            _ => END.to_string(),
        }
    },
    [
        ("research_node", "research_node"),
        ("write_node", "write_node"),
        (END, END),
    ],
)

Routing From Inside a Node

A conditional edge fixes its targets when the graph is built. NodeOutput::with_goto does not: a node writes state and names its successors in the same step, and it may name any node in the graph, including one it has no edge to.

use adk_graph::node::NodeOutput;
use serde_json::json;

// The node decides where control goes, from what it just computed.
async fn triage(ctx: &adk_graph::node::NodeContext) -> adk_graph::error::Result<NodeOutput> {
    let amount = ctx.get("amount").and_then(|v| v.as_f64()).unwrap_or(0.0);
    let next = if amount > 10_000.0 { "escalate" } else { "auto_approve" };
    Ok(NodeOutput::new().with_update("risk", json!(next)).with_goto([next]))
}

The frontier a goto produces is checkpointed like any other, so a paused run resumes into the node the goto chose.

Note: use add_conditional_edges when the possible targets are known when you build the graph — the edges then appear in a rendered diagram. Use a goto when the choice belongs to the node.

Router Helpers

Use built-in routers for common patterns:

use adk_graph::edge::Router;

// Route based on a state field value
.conditional_edge("classifier", Router::by_field("sentiment"), [
    ("positive", "positive_handler"),
    ("negative", "negative_handler"),
    ("neutral", "neutral_handler"),
])

// Route based on boolean field
.conditional_edge("check", Router::by_bool("approved"), [
    ("true", "execute"),
    ("false", "reject"),
])

// Limit iterations
.conditional_edge("loop", Router::max_iterations("count", 5), [
    ("continue", "process"),
    ("done", END),
])

Parallel Execution

Multiple edges from a single node execute in parallel:

let agent = GraphAgent::builder("parallel_processor")
    .channels(&["input", "translation", "summary", "analysis"])
    .node(translator_node)
    .node(summarizer_node)
    .node(analyzer_node)
    .node(combiner_node)
    // All three start simultaneously
    .edge(START, "translator")
    .edge(START, "summarizer")
    .edge(START, "analyzer")
    // Wait for all to complete before combining
    .edge("translator", "combiner")
    .edge("summarizer", "combiner")
    .edge("analyzer", "combiner")
    .edge("combiner", END)
    .build()?;

combiner runs once, after all three branches arrive. Nothing in the code above asks for that: a node with more than one incoming direct edge is deferred automatically at compile time. Branches of unequal length therefore join correctly without configuration.

Two details follow from how the count is taken:

Cyclic Graphs (ReAct Pattern)

Build iterative reasoning agents with cycles:

use adk_core::Part;

// Create agent with tools
let reasoner = Arc::new(
    LlmAgentBuilder::new("reasoner")
        .model(model)
        .instruction("Use tools to answer questions. Provide final answer when done.")
        .tool(search_tool)
        .tool(calculator_tool)
        .build()?
);

let reasoner_node = AgentNode::new(reasoner)
    .with_input_mapper(|state| {
        let question = state.get("question").and_then(|v| v.as_str()).unwrap_or("");
        adk_core::Content::new("user").with_text(question)
    })
    .with_output_mapper(|events| {
        let mut updates = std::collections::HashMap::new();
        let mut has_tool_calls = false;
        let mut response = String::new();

        for event in events {
            if let Some(content) = event.content() {
                for part in &content.parts {
                    match part {
                        Part::FunctionCall { name, .. } => {
                            has_tool_calls = true;
                        }
                        Part::Text { text } => {
                            response.push_str(text);
                        }
                        _ => {}
                    }
                }
            }
        }

        updates.insert("has_tool_calls".to_string(), json!(has_tool_calls));
        updates.insert("response".to_string(), json!(response));
        updates
    });

// Build graph with cycle
let react_agent = StateGraph::with_channels(&["question", "has_tool_calls", "response", "iteration"])
    .add_node(reasoner_node)
    .add_node_fn("counter", |ctx| async move {
        let i = ctx.get("iteration").and_then(|v| v.as_i64()).unwrap_or(0);
        Ok(NodeOutput::new().with_update("iteration", json!(i + 1)))
    })
    .add_edge(START, "counter")
    .add_edge("counter", "reasoner")
    .add_conditional_edges(
        "reasoner",
        |state| {
            let has_tools = state.get("has_tool_calls").and_then(|v| v.as_bool()).unwrap_or(false);
            let iteration = state.get("iteration").and_then(|v| v.as_i64()).unwrap_or(0);

            // Safety limit
            if iteration >= 5 { return END.to_string(); }

            if has_tools {
                "counter".to_string()  // Loop back
            } else {
                END.to_string()  // Done
            }
        },
        [("counter", "counter"), (END, END)],
    )
    .compile()?
    .with_recursion_limit(10);

Multi-Agent Supervisor

Route tasks to specialist agents:

// Create supervisor agent
let supervisor = Arc::new(
    LlmAgentBuilder::new("supervisor")
        .model(model.clone())
        .instruction("Route tasks to: researcher, writer, or coder. Reply with agent name only.")
        .build()?
);

let supervisor_node = AgentNode::new(supervisor)
    .with_output_mapper(|events| {
        let mut updates = std::collections::HashMap::new();
        for event in events {
            if let Some(content) = event.content() {
                let text: String = content.parts.iter()
                    .filter_map(|p| p.text())
                    .collect::<Vec<_>>()
                    .join("")
                    .to_lowercase();

                let next = if text.contains("researcher") { "researcher" }
                    else if text.contains("writer") { "writer" }
                    else if text.contains("coder") { "coder" }
                    else { "done" };

                updates.insert("next_agent".to_string(), json!(next));
            }
        }
        updates
    });

// Build supervisor graph
let graph = StateGraph::with_channels(&["task", "next_agent", "research", "content", "code"])
    .add_node(supervisor_node)
    .add_node(researcher_node)
    .add_node(writer_node)
    .add_node(coder_node)
    .add_edge(START, "supervisor")
    .add_conditional_edges(
        "supervisor",
        Router::by_field("next_agent"),
        [
            ("researcher", "researcher"),
            ("writer", "writer"),
            ("coder", "coder"),
            ("done", END),
        ],
    )
    // Agents report back to supervisor
    .add_edge("researcher", "supervisor")
    .add_edge("writer", "supervisor")
    .add_edge("coder", "supervisor")
    .compile()?;

State Management

State Schema with Reducers

Control how state updates are merged:

let schema = StateSchema::builder()
    .channel("current_step")                    // Overwrite (default)
    .list_channel("messages")                   // Append to list
    .channel_with_reducer("count", Reducer::Sum) // Sum values
    .channel_with_reducer("data", Reducer::Custom(Arc::new(|old, new| {
        // Custom merge logic
        merge_json(old, new)
    })))
    .build();

let agent = GraphAgent::builder("stateful")
    .state_schema(schema)
    // ... nodes and edges
    .build()?;

Reducer Types

Update order

Nodes in one super-step run concurrently and finish in whatever order their work takes. Their state updates are applied in node-name order, not in the order the nodes finished.

The order matters whenever a reducer is not commutative. Append builds an array, so the order is the result; a Custom reducer may be order-sensitive too. Sorting by node name makes a run reproducible: the same graph and the same input give the same state, whatever the timing of a slow dependency.

Where several channels are written by one node, they are applied in channel-name order. Channels do not interact, so this matters only for reading a trace.

Checkpointing

Enable persistent state for fault tolerance and human-in-the-loop:

In-Memory (Development)

use adk_graph::checkpoint::MemoryCheckpointer;

let checkpointer = Arc::new(MemoryCheckpointer::new());

let graph = StateGraph::with_channels(&["task", "result"])
    // ... nodes and edges
    .compile()?
    .with_checkpointer_arc(checkpointer.clone());

SQLite (Production)

use adk_graph::checkpoint::SqliteCheckpointer;

let checkpointer = SqliteCheckpointer::new("checkpoints.db").await?;

let graph = StateGraph::with_channels(&["task", "result"])
    // ... nodes and edges
    .compile()?
    .with_checkpointer(checkpointer);

What a Checkpoint Records

A checkpoint stores the accumulated state, the step number, and the frontier — the nodes that still have to run. It is written after the frontier advances, so resuming never re-executes a node that already completed and never double-applies its updates. A run that finishes checkpoints an empty frontier, so resuming a completed thread returns the final state rather than restarting the graph.

Two cases deliberately checkpoint the frontier that was executing rather than the next one, because the interrupted node has not produced its updates yet and must run again on resume:

Streamed runs checkpoint on the same schedule as blocking runs, including when an interrupt ends the stream, so a human-in-the-loop pause is resumable in either execution mode.

Checkpoint History (Time Travel)

Reads only. TimeTravelHandle::state_history(from, to) returns the state that was stored at each checkpointed step. It executes nothing — no node runs, no event is regenerated, and no side effect repeats. To re-run from a point in history, use fork_at to branch that checkpoint and invoke the graph on the forked thread. The method was previously named replay and documented as re-executing the graph, which it never did.

Checkpoints also enable durable resume — if a graph execution crashes or the process restarts, execution resumes from the last persisted checkpoint rather than starting over. Use SqliteCheckpointer (the sqlite feature) for crash-safe persistence. MemoryCheckpointer holds checkpoints in the process, so they do not survive a restart. Those two are the backends this crate ships; implement the Checkpointer trait for anything else.

// List all checkpoints for a thread
let checkpoints = checkpointer.list("thread-id").await?;
for cp in checkpoints {
    println!("Step {}: {:?}", cp.step, cp.state.get("status"));
}

// Load a specific checkpoint
if let Some(checkpoint) = checkpointer.load_by_id(&checkpoint_id).await? {
    println!("State at step {}: {:?}", checkpoint.step, checkpoint.state);
}

Human-in-the-Loop

Pause execution for human approval using dynamic interrupts:

use adk_graph::{error::GraphError, node::NodeOutput};

// Planner agent assesses risk
let planner_node = AgentNode::new(planner_agent)
    .with_output_mapper(|events| {
        let mut updates = std::collections::HashMap::new();
        for event in events {
            if let Some(content) = event.content() {
                let text: String = content.parts.iter()
                    .filter_map(|p| p.text())
                    .collect::<Vec<_>>()
                    .join("");

                // Extract risk level from LLM response
                let risk = if text.to_lowercase().contains("risk: high") { "high" }
                    else if text.to_lowercase().contains("risk: medium") { "medium" }
                    else { "low" };

                updates.insert("plan".to_string(), json!(text));
                updates.insert("risk_level".to_string(), json!(risk));
            }
        }
        updates
    });

// Review node with dynamic interrupt
let graph = StateGraph::with_channels(&["task", "plan", "risk_level", "approved", "result"])
    .add_node(planner_node)
    .add_node(executor_node)
    .add_node_fn("review", |ctx| async move {
        let risk = ctx.get("risk_level").and_then(|v| v.as_str()).unwrap_or("low");
        let approved = ctx.get("approved").and_then(|v| v.as_bool());

        // Already approved - continue
        if approved == Some(true) {
            return Ok(NodeOutput::new());
        }

        // High/medium risk - interrupt for approval
        if risk == "high" || risk == "medium" {
            return Ok(NodeOutput::interrupt_with_data(
                &format!("{} RISK: Human approval required", risk.to_uppercase()),
                json!({
                    "plan": ctx.get("plan"),
                    "risk_level": risk,
                    "action": "Set 'approved' to true to continue"
                })
            ));
        }

        // Low risk - auto-approve
        Ok(NodeOutput::new().with_update("approved", json!(true)))
    })
    .add_edge(START, "planner")
    .add_edge("planner", "review")
    .add_edge("review", "executor")
    .add_edge("executor", END)
    .compile()?
    .with_checkpointer_arc(checkpointer.clone());

// Execute - may pause for approval
let thread_id = "task-001";
let result = graph.invoke(input, ExecutionConfig::new(thread_id)).await;

match result {
    Err(GraphError::Interrupted(interrupt)) => {
        println!("*** EXECUTION PAUSED ***");
        println!("Reason: {}", interrupt.interrupt);
        println!("Plan awaiting approval: {:?}", interrupt.state.get("plan"));

        // Human reviews and approves...

        // Update state with approval
        graph.update_state(thread_id, [("approved".to_string(), json!(true))]).await?;

        // Resume execution
        let final_result = graph.invoke(State::new(), ExecutionConfig::new(thread_id)).await?;
        println!("Final result: {:?}", final_result.get("result"));
    }
    Ok(result) => {
        println!("Completed without interrupt: {:?}", result);
    }
    Err(e) => {
        println!("Error: {}", e);
    }
}

Static Interrupts

Use interrupt_before or interrupt_after for mandatory pause points:

let graph = StateGraph::with_channels(&["task", "plan", "result"])
    .add_node(planner_node)
    .add_node(executor_node)
    .add_edge(START, "planner")
    .add_edge("planner", "executor")
    .add_edge("executor", END)
    .compile()?
    .with_interrupt_before(&["executor"]);  // Always pause before execution

Streaming Execution

Stream events as the graph executes:

use futures::StreamExt;
use adk_graph::stream::StreamMode;

let stream = agent.stream(input, config, StreamMode::Updates);

while let Some(event) = stream.next().await {
    match event? {
        StreamEvent::NodeStart(name) => println!("Starting: {}", name),
        StreamEvent::Updates { node, updates } => {
            println!("{} updated state: {:?}", node, updates);
        }
        StreamEvent::NodeEnd(name) => println!("Completed: {}", name),
        StreamEvent::Done(state) => println!("Final state: {:?}", state),
        _ => {}
    }
}

Stream Modes

Messages mode reads tokens from Node::execute_stream as they are produced. Each node runs once per super-step in this mode: the node reports its state updates on the stream as a StreamEvent::Updates event, and the executor applies those rather than executing the node a second time to collect them. This matters most for AgentNode, where a second execution would mean a second billed model call per node.

Important: a custom Node that overrides execute_stream must yield a StreamEvent::Updates event carrying its state updates. Without it the node streams events but contributes no state in Messages mode. The default execute_stream, which wraps execute, does this for you.

Timeout policies apply to the streamed execution itself. For a stream, idle_timeout means no event was produced within the limit.

Subgraphs

A compiled graph runs as a node of another through SubgraphNode. The inner graph keeps its own channels, edges and interrupt gates, and exchanges named channels with its parent.

use adk_graph::subgraph::SubgraphNode;
use std::sync::Arc;

let outer = StateGraph::with_channels(&["document", "size"])
    .add_node(
        SubgraphNode::new("measure_doc", Arc::new(inner))
            .with_input("document", "text")
            .with_output("length", "size"),
    )
    .add_edge(START, "measure_doc")
    .add_edge("measure_doc", END)
    .compile()?;

That last row is the difference worth knowing: both schemas are available before anything runs, so a mapping naming a channel neither side declares cannot reach a run and surface as an absent value. A subgraph that exchanges nothing at all is rejected the same way, because it could not affect its parent.

Resuming a pause inside a subgraph

Nothing extra is needed. Invoking the parent again on the same thread re-enters the subgraph, which finds its own checkpoint on <parent thread>/<node name> and continues from where it stopped. Work the subgraph finished before the pause is not repeated, and a pause several levels down resumes the same way — the message names each level it passed through.

A subgraph that declares an interrupt gate but holds no checkpointer is rejected when the parent compiles: it would re-enter at its first node and pay for its finished work a second time.

Because both graphs hold real checkpointers, this survives a process restart: a fresh set of graph objects sharing only the databases resumes the same run.

Handing control back to the parent

A node inside a subgraph can end its own graph and name a node of the graph that holds it:

Ok(NodeOutput::new()
    .with_update("reason", json!("no confident answer"))
    .with_goto_parent(["escalate"]))

The subgraph finishes and projects its output channels as usual, then the parent continues at escalate rather than following the subgraph node's declared edges. The parent validates the target, because it is the only side that knows its own nodes.

Reliability and Cost Controls

Each of these is off by default, so a graph behaves as it did before you set one.

Per-node retry

A transient failure — a rate limit, a dropped connection — otherwise ends the run.

use adk_graph::retry::{RetryOn, RetryPolicy};
use std::time::Duration;

let graph = graph.with_node_retry(
    "call_model",
    RetryPolicy::new(3)
        .with_initial_delay(Duration::from_millis(500))
        .with_max_delay(Duration::from_secs(8))
        .with_backoff_factor(2.0)
        .with_retry_on(RetryOn::Any),
);

The delay grows by backoff_factor, is capped at max_delay, and then has jitter applied. A node with no policy runs once, so retry stays opt-in. A policy from RetryPolicy::default() allows ten attempts, whose nine sleeps total about 243 seconds — lower max_attempts where a caller is waiting on the answer.

An interrupt is never retried, whatever retry_on says: a pause is not a failure. The attempt count is checkpointed, so a resumed run continues the budget rather than restarting it.

Bounding concurrency

A wide fan-out dispatches its whole frontier at once, which can exhaust a connection pool or trip a provider rate limit.

let graph = graph.with_max_concurrency(4);

Nodes beyond the cap wait for a slot. The admission order is the frontier sorted by name, so it does not depend on timing. Imperative child invocations are outside this budget, because a parent awaits its children while holding its own slot.

Per-node timeouts

A TimeoutPolicy caps a single attempt and, with idle_timeout, how long a node may go without reporting progress. Exceeding either gives GraphError::NodeTimedOut, which a retry policy may then act on.

Invoking a node directly

When the number of sub-tasks comes from state rather than from the graph's shape, a node can invoke another node itself:

use adk_graph::child::RunNodeOptions;

let output = ctx
    .run_node_with("reviewer", json!({ "aspect": aspect }), RunNodeOptions::with_run_id(aspect))
    .await?;

The target needs no edge. Each completed child is recorded under <parent>/<child>@<run_id>, so a resumed run returns the recorded answer instead of executing the child again — which matters when the child costs a model call.

Node caching

cache_policy on a node keys its result by node name and current state, with an optional TTL, so an unchanged input skips the work. Requires the node-cache feature; a Redis-backed store is available behind redis-cache.

Delta checkpoints

The delta feature stores the difference between super-steps rather than the whole state, which matters when state is large and steps are many.

Time travel

With the time-travel feature, graph.time_travel(thread_id)? returns a handle over a thread's checkpoint history: list the steps, read the state at one, or fork_at a checkpoint to branch a new thread from it. The call returns Result because every operation reads checkpoints, so a graph with no checkpointer reports GraphError::CheckpointError rather than panicking.

Graph-wide defaults

Repeating the same retry across twenty nodes is easy to get wrong by omission.

use adk_graph::graph::NodeDefaults;

let graph = graph
    .with_node_defaults(NodeDefaults::new().with_retry(RetryPolicy::new(3)))
    .with_node_retry("critical", RetryPolicy::new(10));

A per-node value always wins. NodeDefaults also carries a timeout and a failure handler.

Recovering from a node failure

Once a node's retry budget is spent, a handler may record what happened and name a recovery node instead of ending the run:

let graph = graph.with_node_error_handler("charge", |node, error, _state| {
    Ok(NodeOutput::new()
        .with_update("status", json!(format!("{node} failed: {error}")))
        .with_goto(["compensate"]))
});

Returning Err ends the run as before. An interrupt never reaches a handler, because a pause is not a failure.

Bounding checkpoint growth

A thread accumulates one checkpoint per super-step. A run that lives for days therefore grows without bound, which costs storage and slows list.

use adk_graph::checkpoint::RetentionPolicy;
use std::time::Duration;

let graph = graph
    .with_checkpoint_retention(
        RetentionPolicy::keep_last(50).with_max_age(Duration::from_secs(7 * 24 * 3600)),
    );

Pruning happens after each save, so the cost stays proportional to the run and no external job is needed. The newest checkpoint is never discarded, whatever the policy says, because it is the one a resume loads — keep_last(0) is raised to one, and a thread whose every checkpoint is past the age limit keeps one.

Off by default, so an existing thread keeps its whole history and time travel can still reach every step. Set a policy when a thread is long-lived and you do not need to rewind far.

Rejecting undeclared channels

A channel the schema does not declare takes the overwrite reducer, because that is the fallback for an unknown name. A graph that declared a list channel and then wrote a near-miss name keeps only the last value and reports nothing.

let graph = graph.with_strict_channels();

A node writing an undeclared channel then fails the run with GraphError::UndeclaredChannel, naming the node and the channel. Off by default, and inert when a graph declares no channels at all.

ADK Integration

GraphAgent implements the ADK Agent trait, so it works with:

• Runner: Use with adk-runner for standard execution
• Callbacks: Full support for before/after callbacks
• Sessions: Works with adk-session for conversation history
• Streaming: Returns ADK EventStream

use adk_runner::Runner;

let graph_agent = GraphAgent::builder("workflow")
    .before_agent_callback(|ctx| async {
        println!("Starting graph execution for session: {}", ctx.session_id());
        Ok(())
    })
    .after_agent_callback(|ctx, event| async {
        if let Some(content) = event.content() {
            println!("Graph completed with content");
        }
        Ok(())
    })
    // ... graph definition
    .build()?;

// GraphAgent implements Agent trait - use with Launcher or Runner
// See adk-runner README for Runner configuration

Examples

Validated graph examples in this repository:

cargo run --manifest-path examples/tier_examples/standard/Cargo.toml --bin 11-standard-graph
cargo run --manifest-path examples/tier_examples/standard/Cargo.toml --bin 12-standard-sequential
cargo run --manifest-path examples/competitive_graph_resume/Cargo.toml

The full graph gallery with real LLM integration lives in adk-playground.

Comparison with LangGraph

Two differences are worth stating plainly. run_node_with returns the child's result inline and records it, so a resumed run does not pay for a completed child again, whereas Send hands work to the scheduler and collects it through a reducer. And a hop from a subgraph into its parent has no equivalent here yet.

---

Previous: ← Multi-Agent Systems | Next: Realtime Agents →