Skip to main content

Agent Delegation

Agent delegation enables agents to use other agents as tools, creating specialized workflows where orchestrator agents delegate tasks to specialist agents.

Overview​

How It Works​

  1. Agent Discovery: The system automatically discovers agents from your agent files
  2. Tool Conversion: Each agent becomes a callable tool with its ID as the tool name
  3. Execution: The system routes agent tool calls to the appropriate specialist

Required Configuration​

For an agent to be used as a tool, it needs:

---
id: "specialist-agent"
title: "Specialist Agent"
description: "What this agent specializes in"
---

Critical Requirements:

  • id: Unique identifier that becomes the tool name
  • description: A comprehensive description of the tool's purpose, capabilities, and use cases. This is crucial as the agent relies on it to decide when to invoke the tool - better descriptions lead to more accurate tool selection.

Agent Tool Usage​

When calling another agent as a tool, use this format:

{
  "name": "specialist-agent",
  "arguments": {
    "tasks": ["Specific task description for the agent to perform"]
  }
}

Implementation Example​

Code Review Orchestrator​

Create an orchestrator agent that delegates to specialist agents:

1. Security Specialist (~/.forge/agents/security-specialist.md):

---
id: "security-specialist"
title: "Security Specialist"
description: "Reviews code for security vulnerabilities"
tool_supported: true
tools:
  - read
  - search
max_turns: 30
---
You are a security expert focused on identifying vulnerabilities in code.
Focus on authentication, authorization, and input validation issues.

2. Code Review Orchestrator (./.forge/agents/code-review-orchestrator.md):

---
id: "code-review-orchestrator"
title: "Code Review Orchestrator"
description: "Coordinates code review using specialist agents"
tool_supported: false
tools:
  - security-specialist
  - read
  - write
  - patch
max_turns: 75
---

You are a code review orchestrator that coordinates specialist agents.

## Process:
1. Review the code changes
2. Use the security-specialist agent to check for vulnerabilities
3. Provide comprehensive feedback based on the specialist's analysis

## Usage:
When you need security analysis, call the security-specialist agent with specific tasks:
{
  "name": "security-specialist",
  "arguments": {
    "tasks": [
      "Review authentication code for security issues",
      "Check input validation in user endpoints"
    ]
  }
}

Agent Discovery​

The system loads agents from three sources with precedence:

  1. CWD Agents: ./.forge/agents/*.md (highest priority)
  2. Global Agents: ~/.forge/agents/*.md (Unix/macOS) or %USERPROFILE%\.forge\agents\*.md (Windows)
  3. Built-in Agents: Embedded in application (lowest priority)

CWD agents override global agents with the same ID.