Heldinhow OpenCode Subagents
Overview
OpenCode subagents are specialized AI agents that can be defined and invoked for focused, domain-specific tasks. This skill should be invoked when you need specialized agents with clean context for focused execution, such as code review specialists, testing agents, or documentation agents.
Core Principles
- Single Responsibility: Each subagent should have a clear, focused purpose
- Clean Context: Subagents receive focused context, avoiding pollution from unrelated tasks
- Skill Isolation: Subagents carry their own skills and tools, independent of parent session
- Composition: Multiple subagents can work in parallel on independent tasks
Preparation Checklist
- Define the subagent's purpose and scope of expertise
- Identify the skills the subagent needs to carry
- Determine which tools the subagent should have access to
- Choose appropriate model preferences if different from default
Step-by-Step Process
- Define Agent: Create agent definition in AGENTS.md or YAML file
- Specify Expertise: Define the agent's focus areas and knowledge domains
- Configure Tools: Specify which tools the agent can use
- Set Model: Choose model preferences (temperature, max tokens)
- Invoke: Use @mention or CLI to call the subagent for tasks
Do's and Don'ts
- ✅ Do keep subagent definitions focused - one clear purpose
- ✅ Do provide relevant skills that match the subagent's domain
- ✅ Do use subagents for parallel execution of independent tasks
- ❌ Don't create catch-all agents that try to do everything
- ❌ Don't over-configure - simple is often better
- ❌ Don't forget to verify subagent outputs - assume they can make mistakes
Anti-Patterns
- The Generalist Agent: Creating an agent with no clear focus
- Tool Overload: Giving agents access to too many unrelated tools
- No Verification: Trusting subagent outputs without review
- Context Leakage: Not utilizing proper session isolation when needed