using-subagents

Orchestrates sub-agents for multi-step coding work with delegation contracts and review loops.

Updated Mar 28, 2026
One-click install
npx skills add https://github.com/thedutchvisiongroup/agent-skills --skill using-subagents-thedutchvisiongroup
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: using-subagents
Source: https://github.com/thedutchvisiongroup/agent-skills/tree/main/skills/using-subagents
Command: npx skills add https://github.com/thedutchvisiongroup/agent-skills --skill using-subagents-thedutchvisiongroup

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Delegating coding work to sub-agents often produces duplicated effort, conflicting decisions, unreviewed code, and runaway token costs. This Skill imposes a disciplined six-phase orchestration process so multi-agent work stays controlled, verified, and cost-aware. ## Core Features & Use Cases - Delegation Gate: Decides upfront whether to do the work directly, dispatch one sub-agent, or enter full orchestrator mode where the main agent never writes code. - Approved Work Plans: Requires an explicit user-approved plan with decomposition, agent assignment, quality loops, abort criteria, and a confirmed storage location before any dispatch. - Delegation Contracts & Status Protocol: Ships templates for implementer, reviewer, and explorer sub-agents with a mandatory DONE / DONE_WITH_CONCERNS / BLOCKED / NEEDS_CONTEXT status contract plus report files. - Quality Loops: Enforces independent code review after every implementation, conditional security review, and a maximum of three fix iterations before escalation. - Use Case: When asked to add OAuth login with Google and GitHub providers, the Skill decomposes the work into a shared config task, two parallel-safe provider adapters, and a UI integration task, each dispatched with a contract and independently reviewed. ## Quick Start Ask the agent to implement a multi-file feature using the using-subagents skill and approve the work plan it presents before dispatching.

Frequently Asked Questions about using-subagents

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I orchestrate sub-agents for a large coding task?

Pass the Delegation Gate first to choose between direct work, one sub-agent, or orchestrator mode. Then inventory available agents, write a work plan with decomposition and quality loops, get explicit user approval, and dispatch each task with a delegation contract.

When should I use sub-agents instead of doing the work directly?

Delegate when the task decomposes into bounded independent parts, when verbose intermediate output would flood the main context, or when independent review matters. Do trivial edits, tightly coupled changes, and ambiguous high-stakes work directly.

Does this sub-agent workflow work with Claude Code and OpenCode?

Yes, the skill is harness-agnostic and includes harness notes mapping its concepts to OpenCode and Claude Code mechanisms. It covers dispatch tools, agent discovery directories, nesting depth limits, question-tool behavior, and model routing for each harness.

How do I run sub-agents in parallel without conflicts?

Parallel writing is safe only for provably disjoint scopes: different files, no shared state, no output dependencies, and no pending shared decisions. Read-only explorers are always parallel-safe, and fan-out should be capped at three to five concurrent agents.

What happens when a sub-agent gets blocked or fails review?

A BLOCKED status triggers re-planning: add context, shrink scope, use a stronger model, or escalate to the user. Failed reviews enter a fix loop with the implementer, capped at three iterations before mandatory user escalation.

Can sub-agents dispatch their own sub-agents?

Nesting is limited to depth two and only for approved advisory handoffs, such as a code reviewer handing security suspicions to a security specialist. Implementation nesting is forbidden because it compounds errors, cost, and loss of orchestrator control.