Use this only if your runtime supports subagents/agent delegation.
Personal OS does not require subagents to work. The core architecture is:
AGENTS.md+ wrapper files (CLAUDE.md,CODEX.md,PI.md,OPENCLAW.md)- Skills in
.agents/skills/*/SKILL.md - Workflows in
Workflows/*.md - State and context in
Tasks/,Knowledge/,GOALS.md,BACKLOG.md
Use subagents when a task benefits from strict specialization, for example:
- one agent to synthesize research evidence
- one agent to produce a structured plan
- one agent to verify output quality
Do not use subagents for simple tasks where one agent can complete the work directly.
If your runtime supports delegation:
- Main agent reads
AGENTS.mdand selects the relevant workflow. - Delegate only scoped subtasks (research extraction, plan drafting, verification).
- Require each subagent to reference the same source-of-truth files.
- Main agent merges outputs and applies final verification discipline.
Role: [specialist role]
Objective:
- [single, clear outcome]
Scope:
- Read only:
- AGENTS.md
- Workflows/[workflow].md
- [specific files]
Constraints:
- Follow task schema in AGENTS.md
- Do not modify unrelated files
- Return concise output with explicit assumptions
Verification:
- List checks run
- Mark unknowns clearlyMain prompt:
Use daily-standup workflow and propose top 3 tasks.
If needed, delegate:
1) one subagent for task-state extraction from Tasks/*.md
2) one subagent for goal alignment checks against GOALS.md
Then synthesize and return one final prioritized list.
- Subagent features are runtime-dependent and may differ across Claude Code, Codex, Pi, OpenClaw, and other tools.
- Keep Personal OS portable by storing durable logic in:
.agents/skills/Workflows/AGENTS.md
Back to: Tutorials Home