---
name: multi-agent-router
description: A drop-in multi-agent routing setup for Claude Code — one prompt in, a router dispatches it to specialist subagents (research / build / write) running in parallel, then merges the result. Config + the exact .claude files. Triggers: "multi-agent routing", "one prompt many agents", "claude subagent router", "route to specialists".
---
This is the real routing setup I run in Claude Code— not pseudocode. You drop three files into your project's.claude/
folder and one prompt gets read, dispatched to the right specialists in parallel, and merged back into one result. Given away free by[@buildwith.conrad].
Most people drive Claude one prompt at a time — ask, wait, ask again. But Claude Code can spawn subagents and run them in parallel. This setup adds a router: you give it one instruction, it decides which specialists the job needs (research the facts, build the code, write the copy), fires them at the same time, and stitches the outputs into one answer. You stop babysitting; the routing is done for you.
your one prompt
│
┌───▼───┐
│ router │ reads intent → picks specialists
└─┬─┬─┬─┘
┌───┘ │ └───┐
┌────▼─┐ ┌─▼───┐ ┌▼─────┐
│research│ │build│ │write │ run in parallel
└────┬─┘ └─┬───┘ └┬─────┘
└───┐ │ ┌───┘
┌─▼─▼─▼─┐
│ result │ merged into one answer
└───────┘
Create these under your project root. That's the whole setup.
---
name: researcher
description: Gathers facts, reads the codebase/web, returns a tight findings brief. No code, no prose.
tools: Read, Grep, Glob, WebSearch, WebFetch
model: sonnet
---
You are the research specialist. Given a task, collect only the facts needed to complete it —
relevant files, APIs, constraints, prior art. Return a compact bullet brief. Do not build or write copy.
---
name: builder
description: Writes/edits the actual code for a scoped task. Returns a diff or the files changed.
tools: Read, Edit, Write, Bash, Grep, Glob
model: sonnet
---
You are the build specialist. Given a scoped task + the researcher's brief, implement it cleanly,
matching existing conventions. Return what you changed and why. Do not do unrelated refactors.
---
name: writer
description: Produces the human-facing copy — docs, UI text, release notes, captions.
tools: Read, Grep, Glob
model: sonnet
---
You are the writing specialist. Given the task + context, produce clear, on-brand copy only.
No code. Match the voice of the surrounding material.
---
description: Read one request, dispatch to the right specialist subagents in parallel, merge the result.
---
You are the ROUTER. Read the user's request below and:
1. Decide which specialists it needs (researcher / builder / writer — one or several).
2. Launch the chosen agents **in parallel** with the Task tool, each with a tight, scoped prompt
and the shared context they need.
3. When they return, merge their outputs into ONE coherent result for the user — no raw dumps.
Only involve specialists the request actually needs. Request:
$ARGUMENTS
/route build a /pricing page: research our current plans, build the React page, write the headline + CTA
Claude reads the request, fires researcher
-
builder -
writer
at once, and hands you one
finished result instead of three round-trips. Add your own specialists (tester, reviewer, designer)
by dropping more files in .claude/agents/
.
Works in Claude Code out of the box — subagents + the Task tool are built in. No extra install.
This is one slice of how I run whole builds on autopilot — routing, parallel agents, one merged
output. If it was useful, the Content-to-Cash and First Product kits turn this kind of setup
into an actual sellable system: buildwith.conrad
bio link.
Free giveaway from @buildwith.conrad — keyword ROUTE.