Zoimback Copilot Dev Core Subagent Custom Agent

name: subagent-custom-agent description: Creates custom agent files (.agent.md) for use as subagents in GitHub Copilot for VS Code. Use this skill when the user wants to define a specialized agent, build a worker agent, configure a subagent with specific tools/models/instructions, or control how agents can be invoked. Trigger for requests like "crea un agente personalizado", "create a subagent for X", "make a worker agent", "define a specialized agent", "I need an agent that only does Y", "configure agent invocation", "create a reusable agent component", "agente subagente", or any request to build a custom AI agent component for delegation. Use this skill even when the user just mentions creating agents with restricted tools or specific roles.

Custom Agent Creator for Subagents

Custom agents are .agent.md files that define specialized behavior for use as subagents in GitHub Copilot. Each agent runs in its own context window with specific tools, model, and instructions.

File Location

Custom agents live in .github/copilot-agents/ as .agent.md files:

.github/copilot-agents/
├── planner.agent.md
├── implementer.agent.md
├── security-reviewer.agent.md
└── research-analyst.agent.md

Frontmatter Reference

---
name: Agent Name                      # Required: display name in Copilot UI
description: What this agent does     # Recommended: helps AI select the right agent
tools: ['read', 'search', 'edit']     # Tools available to this agent
model: gpt-4o                         # Optional: override the default model
user-invocable: true                  # Default true; set false for subagent-only agents
disable-model-invocation: false       # Default false; set true to prevent AI from using this as subagent
agents: ['Agent1', 'Agent2']          # Restrict which subagents this agent can invoke (* = all, [] = none)
---

Available Tools

Principle of least privilege: assign only the tools required for the agent's job. Read-only roles (planners, reviewers, researchers) should never have edit, create, or run.

Invocation Control

Note: explicitly listing an agent in the coordinator's agents array overrides disable-model-invocation: true on that worker agent.

Writing the Agent Body

The markdown body is the system prompt. Structure it as:

1. Role statement — what the agent is and what it does
2. Scope definition — what it should and shouldn't handle
3. Output format — what it returns to the main agent (keep it concise)
4. Constraints — any safety, scope, or performance limits

The body should be concise. Subagents receive only the task prompt — they don't inherit main agent context or instructions.

Ready-to-Use Templates

Read-Only Research Agent

---
name: Research Analyst
user-invocable: false
tools: ['read', 'search']
---
You are a research analyst. Given a topic or question, search the codebase and 
documentation for relevant context. Return a concise, structured summary of your 
findings — focus only on what is directly relevant. Do not modify any files.

Code Implementer

---
name: Implementer
user-invocable: false
tools: ['read', 'search', 'edit', 'create']
---
You are a code implementer. You receive a specific task with clear requirements. 
Write clean, idiomatic code that follows the patterns in the existing codebase. 
Confirm what you implemented and any decisions made.

Security Reviewer

---
name: Security Reviewer
user-invocable: false
tools: ['read', 'search']
---
You are a security-focused code reviewer specializing in OWASP Top 10 vulnerabilities.
Review the provided code for: injection risks (SQL, XSS, command injection), broken 
access control, cryptographic failures, insecure deserialization, and sensitive data 
exposure. Return a prioritized list of issues with severity (Critical/High/Medium/Low) 
and a suggested fix for each.

Planner (Read-Only)

---
name: Planner
user-invocable: false
tools: ['read', 'search']
---
Break down complex feature requests into clear, ordered implementation tasks. 
Each task must be self-contained and assignable to a single implementer. 
Incorporate any architectural feedback before finalizing the plan.

Design Principles

• Narrow scope = better results: agents with a single, clear responsibility outperform generalist ones
• Stateless by design: each invocation gets a clean context; don't rely on state from previous subagent calls
• Concise output matters: instruct agents to return summaries, not full transcripts, to keep the main agent's context lean
• Match tools to role: mismatched tools (e.g., edit on a reviewer) introduce risk and noise

Source: Zoimback/copilot-dev-core — distributed by TomeVault.