Agency Agent Router: One Lead, Two Specialists

Agency Agent Router is an AI skill that gives you one entry point into The Agency, the open-source collection of specialist AI agents. For any request, it picks one lead specialist and up to two supporting ones, spells out what each needs and owes the next, and pulls their work into a single decision-ready result. It was published to Taku by the person who created the collection.

What The Agency is

The Agency is a large library of agent definitions — more than 230, according to its README — organised into divisions: engineering, design, marketing, sales, product, project management, testing, security, support, finance, game development, research, and more. Each agent is a specialist with its own personality, process, and deliverables. It's one of the most popular agent collections on GitHub.

That size is the strength and the problem. With hundreds of specialists available, the hard part stops being access to expertise and becomes knowing which expert to ask.

What the router does

It picks a lead. One specialist owns the answer, so there's a clear point of view instead of a committee.

It adds at most two supporting specialists. The cap is deliberate. Past a couple of perspectives, more agents mostly add noise and contradictions to reconcile.

It defines every handoff. For each specialist, the router sets out the evidence they need to work from and the deliverable they owe. Handoffs are where multi-agent systems usually lose information, so this is the part that keeps the result coherent.

It synthesises one result. You get a single, decision-ready output rather than three separate opinions to merge yourself.

The Agency's README describes team scenarios — building a startup MVP, launching a marketing campaign — that each assemble several specialists by hand. The router is the constrained, automatic version of that idea, drawn from a curated cross-functional subset of the collection.

What it won't do

• access private systems — your code, CRM, analytics, or accounts
• take external actions such as sending, publishing, or deploying anything
• pretend to have evidence it wasn't given

Every specialist works from what you provide. If a decision depends on data it doesn't have, the right output says so.

How it differs from installing The Agency

When to reach for it

• A problem spans disciplines and you're not sure which specialist to ask
• You want a recommendation, not a panel discussion
• The full roster feels like too much to browse
• You need a quick cross-functional review before a meeting
• You want to try The Agency's approach before installing the whole collection

What's actually inside

One skill. No connectors, no API keys. In the original repository, each agent is its own Markdown file, and the project ships conversion and install scripts so the same agents work across many coding tools.

Setting it up

The GitHub route. For Claude Code, the original offers an install script (./scripts/install.sh --tool claude-code), or you can copy a division's agent files into ~/.claude/agents/ by hand, where Claude Code picks them up as subagents. You then ask for a specialist by name. The repository also generates integration files for Cursor, Codex, Gemini CLI, and other tools, and the project now has its own desktop app, Agency Agents, that browses the roster and installs it into those tools.

The Taku route. Open the router in Taku and describe what you're trying to decide.

The two aren't mutually exclusive. The full collection is the right choice if you want to call on specific specialists directly every day; the router is for getting one good answer without choosing the team yourself.

Who made it

Michael Sitarzewski created The Agency and maintains it on GitHub under the MIT license. Agency Agent Router was also published to the Taku marketplace by Michael Sitarzewski.

You don't need to be technical to use this

You don't need to understand GitHub, Codex, or Claude Code to use this skill. The original collection is built for people who are comfortable with install scripts and agent directories — and a cross-functional question doesn't usually come from someone who wants to learn either.

Taku is a desktop app built for that: find something useful someone already made, open it in your own workspace, and use it with no setup. Browse the free app library for more, or read about AI agent orchestration for how coordinating several agents works in general. If the decision is about organising a project, the AFFiNE Workspace Blueprint turns it into a linked Markdown plan.

Download Taku and bring your next cross-functional question to a small team of specialists. Taku is in Beta, and the Mac app is available now; there's also an experimental, unsigned Windows build for Windows 10 and 11.

FAQ

What does Agency Agent Router do?

It selects one lead specialist and up to two supporting specialists from a curated subset of The Agency, defines the evidence and deliverable for each handoff, and combines their work into one decision-ready result.

What is The Agency?

An open-source collection of more than 230 specialist AI agent definitions, organised into divisions such as engineering, design, marketing, and finance.

Does it connect to my tools or accounts?

No. It doesn't access private systems or take external actions. It works only from the information you provide.

Is it the same as installing The Agency?

No. Installing the collection gives you every specialist to call by name. The router uses a curated subset and chooses a small team for you.

Is it free?

The skill is free. You need an AI tool that can run it. For what Taku costs, see pricing.

Key points

• One entry point into The Agency's collection of specialist AI agents.
• Picks one lead and at most two supporting specialists per request.
• Defines the evidence and deliverable at every handoff, then merges one result.
• No private-system access and no external actions.
• Published on Taku by Michael Sitarzewski, who created The Agency.