Orchestrate Agent Subagents

Goal

Use subagents only when they create real leverage, with clear ownership and safe integration.

Inputs

• Main task shape
• Chosen architecture or explicit decision that delegation is warranted
• Available agent capabilities
• Shared context and integration constraints

Non-Goals

• Making the top-level single-agent vs multi-agent architecture choice
• Using subagents for status or novelty
• Delegating urgent blocking work without a clear handoff

Workflow

Trigger signals

• Architecture calls for multiple agents but ownership boundaries are not drawn
• User asks "怎么分工" or "how should we split the work"
• Subagents are being launched without clear context contracts or merge rules

1. Restate the delegation boundary

Confirm what part of the system or workflow is already designated for delegation and what must remain local to the coordinating agent. Success criteria: The orchestration work starts from an explicit delegation decision instead of reopening the architecture choice.

2. Split the work into owned units

Assign each subagent a distinct responsibility, output artifact, and write scope. Success criteria: Ownership boundaries prevent duplicated work and merge conflicts.

3. Define context and handoff contracts

Specify what context each subagent receives, what it must return, and what it must not assume. Success criteria: The delegation plan has minimal but sufficient context and clear artifact expectations.

4. Freeze permission, approval, and side-effect inheritance

State which parent-agent permission limits, approval requirements, write scopes, and side-effect restrictions are inherited by each subagent. Specify whether a subagent may execute side effects directly or may only return proposals for the coordinating agent to act on. Success criteria: Delegation does not silently widen autonomy, tool access, or side-effect authority.

5. Define synchronization and merge rules

State when results are needed, how conflicts are resolved, and what main-agent review is required before integration. Success criteria: The orchestration includes a predictable reintegration path instead of ad hoc waiting and trust.

6. Define interruption and no-delegation rules

Record what kinds of tasks should stay local, what should interrupt a running subagent, and when delegation should be abandoned. Success criteria: The system knows when not to delegate or when to take work back.

Output Contract

A delegation plan that includes:

• the delegated scope,
• per-agent ownership,
• context and artifact contracts,
• inherited permission and approval boundaries,
• merge and review rules,
• explicit no-delegation boundaries.

Escalation

Pause when:

• subtasks are too interdependent,
• ownership is blurry,
• a subagent would receive broader tool access or side-effect authority than the

parent agent,

• the very next local step is blocked on delegated results,
• multiple subagents would need to edit the same area without a clear merge

strategy.

Common Failure Modes

• Delegating because it feels sophisticated
• Giving subagents too much context and too little ownership
• Silently widening autonomy through delegation
• Giving multiple agents the same responsibility
• Reopening the top-level architecture choice inside an orchestration skill
• Trusting subagent output without integration review

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bikeread
• Source: bikeread/promethos
• License: MIT

Install and usage instructions live in the source repository linked above.