Sub-agents

Sub-agents are agents that work under a supervisor agent to handle specific tasks. This architecture allows you to create agent workflows where each sub-agent focuses on a specific domain, coordinated by a supervisor.

Why Use Sub-agents?

Task delegation: Assign specific tasks to agents configured for particular domains (e.g., coding, translation, data analysis)
Workflow orchestration: Build multi-step workflows by delegating tasks to appropriate agents
Code organization: Break down complex problems into smaller components
Modularity: Add or swap agents without disrupting the entire system

Creating and Using Sub-agents

Creating Individual Agents

Create the agents that will serve as sub-agents:

import { Agent } from "@voltagent/core";
import { openai } from "@ai-sdk/openai";

// Create an agent for content creation
const contentCreatorAgent = new Agent({
  name: "ContentCreator",
  instructions: "Creates short text content on requested topics",
  model: openai("gpt-4o-mini"),
});

// Create an agent for formatting
const formatterAgent = new Agent({
  name: "Formatter",
  instructions: "Formats and styles text content",
  model: openai("gpt-4o-mini"),
});

Creating a Supervisor Agent

Pass the agents in the subAgents array during supervisor initialization:

import { Agent } from "@voltagent/core";
import { openai } from "@ai-sdk/openai";

const supervisorAgent = new Agent({
  name: "Supervisor",
  instructions: "Coordinates between content creation and formatting agents",
  model: openai("gpt-4o-mini"),
  subAgents: [contentCreatorAgent, formatterAgent],
});

Advanced Sub-agent Configuration

By default, sub-agents use the streamText method. You can specify different methods like generateText, generateObject, or streamObject with custom schemas and options.

See: Advanced Configuration

Customizing Supervisor Behavior

Supervisor agents use an automatically generated system message that includes guidelines for managing sub-agents. Customize this behavior using the supervisorConfig option.

Default System Message

See the generateSupervisorSystemMessage implementation on GitHub.

Type Safety

The supervisorConfig option is only available when subAgents are provided. TypeScript will prevent you from using supervisorConfig on agents without sub-agents.

Basic Supervisor Configuration

import { Agent } from "@voltagent/core";
import { openai } from "@ai-sdk/openai";

const supervisorAgent = new Agent({
  name: "Content Supervisor",
  instructions: "Coordinate content creation workflow",
  model: openai("gpt-4o-mini"),
  subAgents: [writerAgent, editorAgent],

  supervisorConfig: {
    // Add custom guidelines to the default ones
    customGuidelines: [
      "Always thank the user at the end",
      "Keep responses concise and actionable",
      "Prioritize user experience",
    ],

    // Control whether to include previous agent interactions
    includeAgentsMemory: true, // default: true
  },
});

Stream Event Forwarding Configuration

Control which events from sub-agents are forwarded to the parent stream. By default, only tool-call and tool-result events are forwarded.

const supervisorAgent = new Agent({
  name: "Content Supervisor",
  instructions: "Coordinate content creation workflow",
  model: openai("gpt-4o-mini"),
  subAgents: [writerAgent, editorAgent],

  supervisorConfig: {
    // Configure which sub-agent events to forward
    fullStreamEventForwarding: {
      // Default: ['tool-call', 'tool-result']
      types: ["tool-call", "tool-result", "text-delta", "reasoning", "source", "error", "finish"],
    },
  },
});

Common Configurations:

// Minimal - Only tool events (default)
fullStreamEventForwarding: {
  types: ['tool-call', 'tool-result'],
}

// Text + Tools - Include text generation
fullStreamEventForwarding: {
  types: ['tool-call', 'tool-result', 'text-delta'],
}

// Full visibility - All events including reasoning
fullStreamEventForwarding: {
  types: ['tool-call', 'tool-result', 'text-delta', 'reasoning', 'source', 'error', 'finish'],
}

// Clean tool names - No agent prefix (add prefix manually when consuming events if desired)
fullStreamEventForwarding: {
  types: ['tool-call', 'tool-result'],
}

This configuration balances stream performance and information detail for sub-agent interactions.

Error Handling Configuration

Control how the supervisor handles sub-agent failures.

const supervisorAgent = new Agent({
  name: "Supervisor",
  instructions: "Coordinate between agents",
  model: openai("gpt-4o-mini"),
  subAgents: [dataProcessor, analyzer],

  supervisorConfig: {
    // Control whether stream errors throw exceptions
    throwOnStreamError: false, // default: false

    // Control whether error messages appear in empty responses
    includeErrorInEmptyResponse: true, // default: true
  },
});

Configuration Options

throwOnStreamError (boolean, default: false)

When false: Stream errors are caught and returned as error results with status: "error"
When true: Stream errors throw exceptions that must be caught with try/catch
Set to true to handle errors at a higher level or trigger retry logic

includeErrorInEmptyResponse (boolean, default: true)

When true: Error messages are included in the response when no content was generated
When false: Returns empty string in result, but still marks status as "error"
Set to false to handle error messaging yourself

Common Error Handling Patterns

Default - Graceful Error Handling:

// Errors are returned as results with helpful messages
supervisorConfig: {
  throwOnStreamError: false,
  includeErrorInEmptyResponse: true,
}

// Usage:
const result = await supervisor.streamText("Process data");
// If sub-agent fails:
// result contains error message like "Error in DataProcessor: Stream failed"

Native Retry Support

VoltAgent uses the AI SDK's native retry mechanism (default: 3 attempts). Setting throwOnStreamError: true is useful for custom error handling or logging at a higher level, not for implementing retry logic.

Silent Errors - Custom Messaging:

// Errors don't include automatic messages
supervisorConfig: {
  includeErrorInEmptyResponse: false,
}

// Usage with custom error handling:
const result = await supervisor.streamText("Process data");
for await (const event of result.fullStream) {
  if (event.type === "error") {
    // Provide custom user-friendly error message
    console.log("We're having trouble processing your request. Please try again.");
  }
}

Production Setup - Error Tracking:

supervisorConfig: {
  throwOnStreamError: false, // Don't crash the app
  includeErrorInEmptyResponse: true, // Help with debugging

  // Capture error events for monitoring
  fullStreamEventForwarding: {
    types: ['tool-call', 'tool-result', 'error'],
  },
}

Error Handling in Practice

The supervisor's behavior when a sub-agent encounters an error depends on your configuration:

// Example: Sub-agent fails during stream
const supervisor = new Agent({
  name: "Supervisor",
  subAgents: [unreliableAgent],
  supervisorConfig: {
    throwOnStreamError: false,
    includeErrorInEmptyResponse: true,
  },
});

// The supervisor handles the failure
const response = await supervisor.streamText("Do something risky");

// Check the response
if (response.status === "error") {
  console.log("Sub-agent failed:", response.error);
  // response.result contains: "Error in UnreliableAgent: [error details]"
} else {
  // Process successful response
}

Using with fullStream

When using fullStream, the configuration controls what you receive from sub-agents:

// Stream with full event details
const result = await supervisorAgent.streamText("Create and edit content", {
  fullStream: true,
});

// Process different event types
for await (const event of result.fullStream) {
  switch (event.type) {
    case "tool-call":
      console.log(
        `${event.subAgentName ? `[${event.subAgentName}] ` : ""}Tool called: ${event.data.toolName}`
      );
      break;
    case "tool-result":
      console.log(
        `${event.subAgentName ? `[${event.subAgentName}] ` : ""}Tool result: ${event.data.result}`
      );
      break;
    case "text-delta":
      // Only appears if included in types array
      console.log(`Text: ${event.data}`);
      break;
    case "reasoning":
      // Only appears if included in types array
      console.log(`Reasoning: ${event.data}`);
      break;
  }
}

Filtering Sub-agent Events

Identify which events come from sub-agents by checking for subAgentId and subAgentName properties:

const result = await supervisorAgent.streamText("Create and edit content", {
  fullStream: true,
});

for await (const event of result.fullStream) {
  // Check if this event is from a sub-agent
  if (event.subAgentId && event.subAgentName) {
    console.log(`Event from sub-agent ${event.subAgentName}:`);
    console.log(`  Type: ${event.type}`);
    console.log(`  Data:`, event.data);

    // Filter by specific sub-agent
    if (event.subAgentName === "WriterAgent") {
      // Handle writer agent events specifically
    }
  } else {
    // This is from the supervisor agent itself
    console.log(`Supervisor event: ${event.type}`);
  }
}

This allows you to:

Distinguish between supervisor and sub-agent events
Filter events by specific sub-agent
Apply different handling logic based on the event source

Complete System Message Override

Provide a custom systemMessage to replace the default template:

const supervisorAgent = new Agent({
  name: "Custom Supervisor",
  instructions: "This will be ignored when systemMessage is provided",
  model: openai("gpt-4o-mini"),
  subAgents: [writerAgent, editorAgent],

  supervisorConfig: {
    systemMessage: `
You are a content manager named "ContentBot".

Your team:
- Writer: Creates original content
- Editor: Reviews and improves content

Your workflow:
1. Analyze user requests
2. Use delegate_task to assign work to appropriate specialists
3. Coordinate between specialists as needed
4. Provide final responses
5. Maintain a professional tone

Remember: Use the delegate_task tool to assign tasks to your specialists.
    `.trim(),

    // Control memory inclusion even with custom system message
    includeAgentsMemory: true,
  },
});

Quick Usage

Add custom rules:

supervisorConfig: {
  customGuidelines: ["Verify sources", "Include confidence levels"];
}

Override entire system message:

supervisorConfig: {
  systemMessage: "You are TaskBot. Use delegate_task(task, [agentNames]) to assign work.";
}

Control memory:

supervisorConfig: {
  includeAgentsMemory: false; // Fresh context each interaction (default: true)
}

Configure event forwarding:

supervisorConfig: {
  fullStreamEventForwarding: {
    types: ['tool-call', 'tool-result', 'text-delta'], // Control which events to forward
  }
}

Handle errors gracefully:

supervisorConfig: {
  throwOnStreamError: false, // Return errors as results (default)
  includeErrorInEmptyResponse: true // Include error details in response (default)
}

Throw exceptions for custom error handling:

supervisorConfig: {
  throwOnStreamError: true; // Throw exceptions on sub-agent failures
}

How Sub-agents Work

The supervisor agent delegates tasks to its sub-agents using the automatically provided delegate_task tool.

A user sends a request to the supervisor agent.
The supervisor's LLM analyzes the request and its system prompt (which lists available sub-agents).
Based on the task, the supervisor decides which sub-agent(s) to use.
The supervisor uses the delegate_task tool to hand off the task(s).

The `delegate_task` Tool

This tool is automatically added to supervisor agents and handles delegation.

Name: delegate_task
Description: "Delegate a task to one or more specialized agents"
Parameters:
- task (string, required): The task description to be delegated
- targetAgents (array of strings, required): Sub-agent names to delegate the task to. The supervisor can delegate to multiple agents simultaneously
- context (object, optional): Additional context needed by the sub-agent(s)
Execution:
- Finds the sub-agent instances based on the provided names
- Calls the handoffTask (or handoffToMultiple) method internally
- Passes the supervisor's agent ID (parentAgentId) and history entry ID (parentHistoryEntryId) for observability

Returns:

Always returns an array of result objects (even for single agent):

[
  {
    agentName: string; // Name of the sub-agent that executed the task
    response: string; // The text result returned by the sub-agent
    usage?: UsageInfo; // Token usage information
    bailed?: boolean; // True if onHandoffComplete called bail()
  },
  // ... more results if multiple agents were targeted
]

When bailed: true, the supervisor's execution is terminated immediately and the subagent's response is returned to the user. See Early Termination (Bail) for details.

Sub-agents process their delegated tasks independently. They can use their own tools or delegate further if they are also supervisors.
Each sub-agent returns its result to the delegate_task tool execution context.
The supervisor receives the results from the delegate_task tool.
The supervisor synthesizes the final response based on its instructions and the received results.

Complete Example

import { Agent } from "@voltagent/core";
import { openai } from "@ai-sdk/openai";

// Create agents
const writer = new Agent({
  name: "Writer",
  instructions: "Write creative stories",
  model: openai("gpt-4o-mini"),
});

const translator = new Agent({
  name: "Translator",
  instructions: "Translate text accurately",
  model: openai("gpt-4o-mini"),
});

// Create supervisor
const supervisor = new Agent({
  name: "Supervisor",
  instructions: "Coordinate story writing and translation",
  model: openai("gpt-4o-mini"),
  subAgents: [writer, translator],
});

// Use it
const result = await supervisor.streamText("Write a story about AI and translate to Spanish");

for await (const chunk of result.textStream) {
  process.stdout.write(chunk);
}

What happens:

Supervisor analyzes request
Calls delegate_task → Writer creates story
Calls delegate_task → Translator translates
Combines results and responds

Using Hooks

VoltAgent provides hooks to monitor and control the supervisor/subagent workflow:

`onHandoff` Hook

Triggered when delegation begins:

const supervisor = new Agent({
  name: "Supervisor",
  subAgents: [writer, translator],
  hooks: {
    onHandoff: ({ agent, sourceAgent }) => {
      console.log(`${sourceAgent.name} → ${agent.name}`);
    },
  },
});

`onHandoffComplete` Hook

Triggered when a subagent completes execution. This hook enables early termination (bail) to optimize token usage:

const supervisor = new Agent({
  name: "Supervisor",
  subAgents: [dataAnalyzer, reportGenerator],
  hooks: {
    onHandoffComplete: async ({ agent, sourceAgent, result, messages, usage, context, bail }) => {
      // Bail if subagent produced final output
      if (agent.name === "Report Generator") {
        context.logger?.info("Final report ready, bailing");
        bail(); // Skip supervisor processing
      }
    },
  },
});

See Early Termination (Bail) below for detailed usage.

Sub-agents automatically inherit the supervisor's context:

// Supervisor passes context
const response = await supervisor.streamText("Task", {
  context: new Map([["projectId", "123"]]),
});

// Sub-agent receives it automatically
const subAgent = new Agent({
  hooks: {
    onStart: (context) => {
      const projectId = context.context.get("projectId"); // "123"
    },
  },
});

Early Termination (Bail)

The Problem

In supervisor/subagent workflows, subagents always return to the supervisor for processing, even when they generate final outputs (like JSON structures or reports) that need no additional handling. This wastes tokens:

Current flow:
Supervisor → SubAgent (generates 2K token JSON) → Supervisor (processes JSON) → User
                                                    ↑ Wastes ~2K tokens

Example impact:

Without bail: ~2,650 tokens per request
With bail: ~560 tokens per request
Savings: 79% (~$0.020 per request)

The Solution

The onHandoffComplete hook allows supervisors to intercept subagent results and bail (skip supervisor processing) when the subagent produces final output:

New flow:
Supervisor → SubAgent → bail() → User ✅

Basic Usage

Call bail() in the onHandoffComplete hook to terminate early:

const supervisor = new Agent({
  name: "Workout Supervisor",
  subAgents: [exerciseAgent, workoutBuilder],
  hooks: {
    onHandoffComplete: async ({ agent, result, bail, context }) => {
      // Workout Builder produces final JSON - no processing needed
      if (agent.name === "Workout Builder") {
        context.logger?.info("Final output received, bailing");
        bail(); // Skip supervisor, return directly to user
      }
      // Default: continue to supervisor for processing
    },
  },
});

Conditional Bail Logic

Bail based on agent name, result size, or content:

hooks: {
  onHandoffComplete: async ({ agent, result, bail, context }) => {
    // By agent name
    if (agent.name === "Report Generator") {
      bail();
      return;
    }

    // By result size (save tokens)
    if (result.length > 2000) {
      context.logger?.warn("Large result, bailing to save tokens");
      bail();
      return;
    }

    // By result content
    if (result.includes("FINAL_OUTPUT")) {
      bail();
      return;
    }

    // Default: continue to supervisor
  },
}

Transform Before Bail

Optionally transform the result before bailing:

hooks: {
  onHandoffComplete: async ({ agent, result, bail }) => {
    if (agent.name === "Report Generator") {
      // Add metadata before returning
      const transformed = `# Final Report\n\n${result}\n\n---\nGenerated at: ${new Date().toISOString()}`;
      bail(transformed); // Bail with transformed result
    }
  },
}

Hook Parameters

interface OnHandoffCompleteHookArgs {
  agent: Agent; // Target agent (subagent)
  sourceAgent: Agent; // Source agent (supervisor)
  result: string; // Subagent's output
  messages: UIMessage[]; // Full conversation messages
  usage?: UsageInfo; // Token usage info
  context: OperationContext; // Operation context
  bail: (transformedResult?: string) => void; // Call to bail
}

Accessing Bailed Results

When a subagent bails, the subagent's result is returned to the user (not the supervisor's):

const supervisor = new Agent({
  name: "Supervisor",
  subAgents: [
    createSubagent({
      agent: workoutBuilder,
      method: "generateObject",
      schema: WorkoutSchema,
    }),
  ],
  hooks: {
    onHandoffComplete: async ({ agent, bail }) => {
      if (agent.name === "Workout Builder") {
        bail(); // Return workout JSON directly
      }
    },
  },
});

const result = await supervisor.generateText("Create workout");
console.log(result.text); // Contains workout JSON, not supervisor's processing

Supported Methods

Bail works with methods that support tools:

✅ generateText - Aborts execution, returns bailed result
✅ streamText - Aborts stream immediately, returns bailed result
❌ generateObject - No tool support, bail not applicable
❌ streamObject - No tool support, bail not applicable

Stream Event Visibility with Bail

When using bail with toUIMessageStream() or consuming fullStream events, you need to configure which events are forwarded from subagents.

Default Behavior:

By default, only tool-call and tool-result events are forwarded from subagents. This means subagent text chunks are NOT visible in the stream:

// Default configuration (implicit)
supervisorConfig: {
  fullStreamEventForwarding: {
    types: ['tool-call', 'tool-result'], // ⚠️ text-delta NOT included
  }
}

To See Subagent Text Output:

When a subagent bails and produces text output, you must explicitly include text-delta in the forwarded event types:

const supervisor = new Agent({
  name: "Supervisor",
  subAgents: [workoutBuilder],
  supervisorConfig: {
    fullStreamEventForwarding: {
      types: ["tool-call", "tool-result", "text-delta"], // ✅ Include text-delta
    },
  },
  hooks: {
    onHandoffComplete: ({ agent, bail }) => {
      if (agent.name === "Workout Builder") {
        bail(); // Subagent text will be visible in stream
      }
    },
  },
});

Consuming Bailed Subagent Output:

When using toUIMessageStream(), subagent text appears as data-subagent-stream events that are automatically grouped and rendered in the UI:

const result = await supervisor.streamText("Create workout");

// With text-delta forwarding enabled:
// Subagent text chunks are accumulated and displayed as a collapsible box
// with the subagent's name (e.g., "Workout Builder")

for await (const message of result.toUIMessageStream()) {
  // Message parts include grouped subagent output
  // Rendered automatically in observability UI
}

Event Types Reference:

Event Type	Description	Default
`tool-call`	Tool invocations	✅ Included
`tool-result`	Tool results	✅ Included
`text-delta`	Text chunk generation	❌ NOT included
`reasoning`	Model reasoning (if available)	❌ NOT included
`source`	Retrieved sources (if available)	❌ NOT included
`error`	Error events	❌ NOT included
`finish`	Stream completion	❌ NOT included

Observability

Bailed subagents are tracked in observability with visual indicators:

Logging:

[INFO] Supervisor bailed after handoff
  supervisor: Workout Supervisor
  subagent: Workout Builder
  transformed: false
  resultLength: 450

OpenTelemetry Attributes:

Both supervisor and subagent spans get attributes:

// Supervisor span attributes
{
  "bailed": true,
  "bail.subagent": "Workout Builder",
  "bail.transformed": false
}

// Subagent span attributes
{
  "bailed": true,
  "bail.supervisor": "Workout Supervisor",
  "bail.transformed": false
}

Use Cases

Perfect for scenarios where specialized subagents generate final outputs:

JSON/Structured data generators - Workout builders, report generators, data exporters
Large content producers - Document creators, extensive data analysis
Token optimization - Skip processing for expensive results
Business logic - Conditional routing based on result characteristics

Best Practices

✅ DO:

Bail when subagent produces final, ready-to-use output
Use conditional logic to bail selectively
Log bail decisions for debugging
Transform results before bailing when needed

❌ DON'T:

Bail on every subagent (defeats supervisor purpose)
Bail when supervisor needs to process/combine results
Forget to handle non-bailed flow (default case)

## Step Control

Control workflow steps with `maxSteps`:

```ts
const supervisor = new Agent({
  subAgents: [writer, editor],
  maxSteps: 20, // Inherited by all sub-agents
});

// Override per request
const result = await supervisor.generateText("Task", { maxSteps: 10 });

Default: 10 × number_of_sub-agents (prevents infinite loops)

Observability

Sub-agent operations are automatically linked to their supervisor for traceability in monitoring tools.

Advanced Configuration

Use different execution methods for sub-agents:

import { createSubagent } from "@voltagent/core";
import { z } from "zod";

const AnalysisSchema = z.object({
  insights: z.array(z.string()),
  confidence: z.number().min(0).max(1),
});

const supervisor = new Agent({
  subAgents: [
    writer, // Default streamText
    createSubagent({
      agent: analyzer,
      method: "generateObject",
      schema: AnalysisSchema,
      options: { temperature: 0.1 },
    }),
  ],
});

Available methods:

streamText (default) - Real-time text streaming
generateText - Text generation
generateObject - Structured data with Zod schema
streamObject - Streaming structured data

Dynamic Sub-agents

Add sub-agents after initialization:

supervisor.addSubAgent(newAgent);

Remove Sub-agents

supervisor.removeSubAgent(agentId);

Troubleshooting

Sub-agent not being called?

Check agent names match exactly
Make supervisor instructions explicit about when to delegate
Use onHandoff hook to debug delegation flow

Sub-agents

Why Use Sub-agents?​

Creating and Using Sub-agents​

Creating Individual Agents​

Creating a Supervisor Agent​

Customizing Supervisor Behavior​

Basic Supervisor Configuration​

Stream Event Forwarding Configuration​

Error Handling Configuration​

Configuration Options​

Common Error Handling Patterns​

Error Handling in Practice​

Using with fullStream​

Filtering Sub-agent Events​

Complete System Message Override​

Quick Usage​

How Sub-agents Work​

The delegate_task Tool​

Complete Example​

Using Hooks​

onHandoff Hook​

onHandoffComplete Hook​

Context Sharing​

Early Termination (Bail)​

The Problem​

The Solution​

Basic Usage​

Conditional Bail Logic​

Transform Before Bail​

Hook Parameters​

Accessing Bailed Results​

Supported Methods​

Stream Event Visibility with Bail​

Observability​

Use Cases​

Best Practices​

Observability​

Advanced Configuration​

Dynamic Sub-agents​

Remove Sub-agents​

Troubleshooting​

Table of Contents

Why Use Sub-agents?

Creating and Using Sub-agents

Creating Individual Agents

Creating a Supervisor Agent

Customizing Supervisor Behavior

Basic Supervisor Configuration

Stream Event Forwarding Configuration

Error Handling Configuration

Configuration Options

Common Error Handling Patterns

Error Handling in Practice

Using with fullStream

Filtering Sub-agent Events

Complete System Message Override

Quick Usage

How Sub-agents Work

The `delegate_task` Tool

Complete Example

Using Hooks

`onHandoff` Hook

`onHandoffComplete` Hook

Context Sharing

Early Termination (Bail)

The Problem

The Solution

Basic Usage

Conditional Bail Logic

Transform Before Bail

Hook Parameters

Accessing Bailed Results

Supported Methods

Stream Event Visibility with Bail

Observability

Use Cases

Best Practices

Observability

Advanced Configuration

Dynamic Sub-agents

Remove Sub-agents

Troubleshooting