Workflow Engine

Overview

MoFA’s workflow engine is inspired by LangGraph and provides a stateful, graph-based workflow system for orchestrating complex agent processes. Key Features:

State Management: Reducers for configurable state updates
Control Flow: Conditional edges, loops, branching
Persistence: Checkpoint/restore workflow state
Telemetry: Debug sessions with step-by-step replay
MapReduce: Dynamic edge creation for parallel tasks

Core Concepts

StateGraph

The StateGraph is the main abstraction for building workflows:

use mofa_kernel::workflow::{StateGraph, START, END};
use serde::{Serialize, Deserialize};

#[derive(Clone, Serialize, Deserialize)]
struct MyState {
    messages: Vec<String>,
    step_count: u32,
}

let graph = StateGraphImpl::<MyState>::new("my_workflow")
    .add_node("process", Box::new(ProcessNode))
    .add_node("validate", Box::new(ValidateNode))
    .add_edge(START, "process")
    .add_edge("process", "validate")
    .add_edge("validate", END)
    .compile()?;

Graph State Trait

All workflow states must implement GraphState:

mofa-kernel/src/workflow/state.rs

pub trait GraphState: Clone + Send + Sync + 'static {
    /// Schema for state validation
    fn schema() -> StateSchema;
    
    /// Apply an update to the state
    fn update(&mut self, update: StateUpdate) -> Result<(), WorkflowError>;
    
    /// Serialize to JSON
    fn to_json(&self) -> serde_json::Value;
    
    /// Deserialize from JSON
    fn from_json(value: serde_json::Value) -> Result<Self, WorkflowError>;
}

Node Functions

Nodes are async functions that process state:

use mofa_kernel::workflow::{NodeFunc, RuntimeContext, Command};

struct ProcessNode;

#[async_trait]
impl NodeFunc<MyState> for ProcessNode {
    async fn call(
        &self,
        state: MyState,
        ctx: &RuntimeContext,
    ) -> Result<Command<MyState>, WorkflowError> {
        // Process the state
        let mut new_state = state.clone();
        new_state.messages.push("Processed".to_string());
        new_state.step_count += 1;
        
        // Return update command
        Ok(Command::Update(new_state))
    }
}

Reducer System

Reducers control how state updates are merged:

Overwrite
Append
Prepend
Merge
Custom

Replace the entire field:

use mofa_kernel::workflow::{Reducer, ReducerType};

let reducer = Reducer::new(ReducerType::Overwrite);

// Old state: { messages: ["A", "B"] }
// Update: { messages: ["C"] }
// Result: { messages: ["C"] }

Append to array/list:

let reducer = Reducer::new(ReducerType::Append);

// Old state: { messages: ["A", "B"] }
// Update: { messages: ["C"] }
// Result: { messages: ["A", "B", "C"] }

Prepend to array/list:

let reducer = Reducer::new(ReducerType::Prepend);

// Old state: { messages: ["B", "C"] }
// Update: { messages: ["A"] }
// Result: { messages: ["A", "B", "C"] }

Deep merge objects:

let reducer = Reducer::new(ReducerType::Merge);

// Old state: { user: { name: "Alice", age: 30 } }
// Update: { user: { age: 31, city: "NYC" } }
// Result: { user: { name: "Alice", age: 31, city: "NYC" } }

User-defined merge logic:

let reducer = Reducer::new(ReducerType::Custom(Box::new(
    |old: &serde_json::Value, new: &serde_json::Value| {
        // Custom merge logic
        Ok(new.clone())
    }
)));

Configure per field:

impl GraphState for MyState {
    fn schema() -> StateSchema {
        StateSchema::new()
            .add_field("messages", ReducerType::Append)
            .add_field("step_count", ReducerType::Overwrite)
    }
}

Command Pattern

Nodes return Command objects for control flow:

mofa-kernel/src/workflow/command.rs

pub enum Command<S: GraphState> {
    /// Update state and continue
    Update(S),
    
    /// Update state and specify next node(s)
    UpdateAndGoto {
        state: S,
        next: ControlFlow,
    },
    
    /// End workflow
    End(S),
    
    /// Create dynamic edges (MapReduce)
    Send(SendCommand<S>),
}

pub enum ControlFlow {
    /// Single next node
    Next(String),
    
    /// Multiple parallel nodes
    Parallel(Vec<String>),
    
    /// Conditional routing
    Branch(HashMap<String, bool>),
}

Example: Conditional routing

#[async_trait]
impl NodeFunc<MyState> for RouterNode {
    async fn call(
        &self,
        state: MyState,
        ctx: &RuntimeContext,
    ) -> Result<Command<MyState>, WorkflowError> {
        if state.step_count > 5 {
            Ok(Command::UpdateAndGoto {
                state,
                next: ControlFlow::Next("finalize".to_string()),
            })
        } else {
            Ok(Command::UpdateAndGoto {
                state,
                next: ControlFlow::Next("continue_processing".to_string()),
            })
        }
    }
}

Send Pattern (MapReduce)

Dynamically create edges for parallel task distribution:

use mofa_kernel::workflow::{Command, SendCommand};

#[async_trait]
impl NodeFunc<MyState> for MapNode {
    async fn call(
        &self,
        state: MyState,
        ctx: &RuntimeContext,
    ) -> Result<Command<MyState>, WorkflowError> {
        // Split task into chunks
        let chunks = vec!["chunk1", "chunk2", "chunk3"];
        
        Ok(Command::Send(SendCommand {
            state,
            sends: chunks.into_iter().map(|chunk| {
                Send::new("worker")
                    .with_data("chunk", json!(chunk))
            }).collect(),
            next: Some("reducer".to_string()),
        }))
    }
}

Recursion Limit

Prevent infinite loops with RemainingSteps:

use mofa_kernel::workflow::{GraphConfig, RemainingSteps};

let config = GraphConfig {
    recursion_limit: Some(100),  // Max 100 steps
    ..Default::default()
};

let graph = StateGraphImpl::<MyState>::new("workflow")
    .with_config(config)
    .compile()?;

Behavior:

Each step decrements the counter
When counter reaches 0, workflow stops
Prevents runaway loops

// Inside a node
if ctx.remaining_steps().is_exhausted() {
    return Ok(Command::End(state));
}

Telemetry & Debug Sessions

Record workflow execution for debugging:

use mofa_kernel::workflow::{
    SessionRecorder, DebugSession, DebugEvent
};

let recorder = SessionRecorder::new();
let session_id = recorder.start_session("my-workflow").await;

// Execute workflow with recording
let result = graph.invoke_with_recorder(
    initial_state,
    Some(&recorder),
).await?;

// Replay session
let session = recorder.get_session(&session_id).await?;
for event in session.events() {
    match event {
        DebugEvent::NodeEnter { node, state, .. } => {
            println!("Entered: {}, State: {:?}", node, state);
        }
        DebugEvent::NodeExit { node, result, .. } => {
            println!("Exited: {}, Result: {:?}", node, result);
        }
        _ => {}
    }
}

Recorded Events:

NodeEnter: Node execution started
NodeExit: Node execution completed
StateUpdate: State changed
EdgeTraversal: Graph edge followed
Error: Execution error

Streaming Execution

Stream intermediate steps during execution:

let mut stream = graph.stream(initial_state, None).await?;

while let Some(event) = stream.next().await {
    match event {
        StreamEvent::NodeStarted { node } => {
            println!("Starting: {}", node);
        }
        StreamEvent::NodeCompleted { node, state } => {
            println!("Completed: {}", node);
            println!("State: {:?}", state);
        }
        StreamEvent::WorkflowComplete { state } => {
            println!("Done: {:?}", state);
        }
    }
}

Persistence (Checkpointing)

Save and restore workflow state:

use mofa_kernel::workflow::Checkpointer;

// Save checkpoint
let checkpoint = graph.checkpoint(&state)?;
storage.save("checkpoint-1", &checkpoint).await?;

// Restore and resume
let checkpoint = storage.load("checkpoint-1").await?;
let restored_state = graph.restore_checkpoint(checkpoint)?;
let result = graph.invoke(restored_state, None).await?;

Use Cases:

Long-running workflows (resume after crash)
Human-in-the-loop (wait for approval, resume later)
Time travel debugging (restore to previous state)

Workflow DSL

Define workflows in YAML/JSON:

workflow.yaml

name: data_pipeline
version: "1.0"

nodes:
  - id: fetch
    type: script
    script: |
      fn execute(state) {
        state.data = fetch_from_api();
        state
      }
  
  - id: process
    type: script
    script: |
      fn execute(state) {
        state.data = transform(state.data);
        state
      }
  
  - id: save
    type: script
    script: |
      fn execute(state) {
        save_to_db(state.data);
        state
      }

edges:
  - from: __start__
    to: fetch
  
  - from: fetch
    to: process
  
  - from: process
    to: save
  
  - from: save
    to: __end__

Load and execute:

use mofa_foundation::workflow::dsl::WorkflowLoader;

let loader = WorkflowLoader::new();
let graph = loader.from_file("workflow.yaml").await?;
let result = graph.invoke(initial_state, None).await?;

Complete Example

Scenario: Multi-step code generation workflow

use mofa_kernel::workflow::*;
use serde::{Serialize, Deserialize};

#[derive(Clone, Serialize, Deserialize, Debug)]
struct CodeGenState {
    requirement: String,
    design: Option<String>,
    code: Option<String>,
    tests: Option<String>,
    review_passed: bool,
}

impl GraphState for CodeGenState {
    fn schema() -> StateSchema {
        StateSchema::new()
            .add_field("requirement", ReducerType::Overwrite)
            .add_field("design", ReducerType::Overwrite)
            .add_field("code", ReducerType::Overwrite)
            .add_field("tests", ReducerType::Append)
    }
    
    // ... implement other required methods
}

// Nodes
struct DesignNode;
struct CodeNode;
struct TestNode;
struct ReviewNode;

#[async_trait]
impl NodeFunc<CodeGenState> for DesignNode {
    async fn call(
        &self,
        mut state: CodeGenState,
        ctx: &RuntimeContext,
    ) -> Result<Command<CodeGenState>, WorkflowError> {
        // Call LLM to generate design
        state.design = Some("System design document...".to_string());
        Ok(Command::Update(state))
    }
}

#[async_trait]
impl NodeFunc<CodeGenState> for ReviewNode {
    async fn call(
        &self,
        mut state: CodeGenState,
        ctx: &RuntimeContext,
    ) -> Result<Command<CodeGenState>, WorkflowError> {
        // Review the code
        let passed = true;  // Simulate review
        state.review_passed = passed;
        
        if passed {
            Ok(Command::UpdateAndGoto {
                state,
                next: ControlFlow::Next(END.to_string()),
            })
        } else {
            // Loop back to code generation
            Ok(Command::UpdateAndGoto {
                state,
                next: ControlFlow::Next("code".to_string()),
            })
        }
    }
}

// Build workflow
let graph = StateGraphImpl::<CodeGenState>::new("codegen")
    .add_node("design", Box::new(DesignNode))
    .add_node("code", Box::new(CodeNode))
    .add_node("test", Box::new(TestNode))
    .add_node("review", Box::new(ReviewNode))
    .add_edge(START, "design")
    .add_edge("design", "code")
    .add_edge("code", "test")
    .add_edge("test", "review")
    // Review -> END or Review -> code (conditional)
    .compile()?;

// Execute
let initial = CodeGenState {
    requirement: "Build a REST API for user management".to_string(),
    design: None,
    code: None,
    tests: None,
    review_passed: false,
};

let result = graph.invoke(initial, None).await?;
println!("Final state: {:?}", result);

Workflow Visualization

Generate Mermaid diagrams from workflows:

let mermaid = graph.to_mermaid();
println!("{}", mermaid);

Output:

Performance Tips

Minimize State Size

Large states slow serialization. Keep only necessary data.

Use Parallel Edges

When nodes are independent, use ControlFlow::Parallel.

Batch Updates

Update state once per node instead of incrementally.

Async All The Way

Ensure all node functions are truly async for concurrency.

Next Steps

Examples

See workflow examples in action

API Reference

Complete workflow API documentation

Get Started

Core Concepts

Building Agents

Plugin Development

Multi-Agent Systems

Advanced Features

Language Bindings

Workflow Engine

Overview

Core Concepts

StateGraph

Graph State Trait

Node Functions

Reducer System

Command Pattern

Send Pattern (MapReduce)

Recursion Limit

Telemetry & Debug Sessions

Streaming Execution

Persistence (Checkpointing)

Workflow DSL

Complete Example

Workflow Visualization

Performance Tips

Minimize State Size

Use Parallel Edges

Batch Updates

Async All The Way

Next Steps

Examples

API Reference

Build docs developers (and LLMs) love

Get Started

Core Concepts

Building Agents

Plugin Development

Multi-Agent Systems

Advanced Features

Language Bindings

​Overview

​Core Concepts

​StateGraph

​Graph State Trait

​Node Functions

​Reducer System

​Command Pattern

​Send Pattern (MapReduce)

​Recursion Limit

​Telemetry & Debug Sessions

​Streaming Execution

​Persistence (Checkpointing)

​Workflow DSL

​Complete Example

​Workflow Visualization

​Performance Tips

Minimize State Size

Use Parallel Edges

Batch Updates

Async All The Way

​Next Steps

Examples

API Reference

Build docs developers (and LLMs) love

Overview

Core Concepts

StateGraph

Graph State Trait

Node Functions

Reducer System

Command Pattern

Send Pattern (MapReduce)

Recursion Limit

Telemetry & Debug Sessions

Streaming Execution

Persistence (Checkpointing)

Workflow DSL

Complete Example

Workflow Visualization

Performance Tips

Next Steps