Build an Entire AI Department

Difficulty: Advanced · Estimated time: 25 min

What You'll Learn

How to combine several specialised agents into a team that covers a whole function, each with a distinct job and a clean handoff to the next.

Why This Matters

One agent handles one task well. A real function — research, marketing, sales, operations — is usually a chain of tasks, each needing a different specialism. Trying to make one agent do all of it produces a generalist again, with all the inconsistency you were trying to avoid. A department is several narrow agents, each excellent at one stage, passing work along the chain.

The Concept

A department is a set of agents where:

For example, a content department might be a Researcher (gathers and summarises sources), a Writer (turns the research into a draft), and an Editor (checks the draft against style and flags issues) — three agents, three Projects or Skills, each only doing its own stage.

Step-by-Step

Example

A small sales department, three agents:

Lead Qualifier — input: a raw inbound enquiry; output: a short structured summary (company, need, urgency, fit score) that the next agent can act on without re-reading the original enquiry.

Outreach Drafter — input: the Qualifier's structured summary; output: a first-draft outreach email matching the lead's fit score and need.

Follow-Up Agent — input: the outreach sent plus any reply received; output: a recommended next action (follow up, escalate to a human, close out) with a one-line reason.

Each agent only ever sees the structured output of the one before it, not the raw original enquiry — that's the handoff doing its job.

Try This

Take a function you're already building agents for and write its workflow as three to five stages on paper, before assigning any agent. Check: does each stage genuinely need a different specialism, or are two of them really the same job?

Copy This

DEPARTMENT: [FUNCTION NAME, e.g. Content, Sales, Research]

STAGE 1: [STAGE NAME]
Agent: [AGENT NAME]
Input: [WHAT IT RECEIVES]
Output: [WHAT IT PRODUCES, EXACT FORMAT]

STAGE 2: [STAGE NAME]
Agent: [AGENT NAME]
Input: [MUST MATCH STAGE 1's OUTPUT FORMAT]
Output: [WHAT IT PRODUCES, EXACT FORMAT]

STAGE 3: [STAGE NAME]
Agent: [AGENT NAME]
Input: [MUST MATCH STAGE 2's OUTPUT FORMAT]
Output: [FINAL DELIVERABLE]

OWNER OF FULL CHAIN: [YOU, OR WHOEVER REVIEWS END-TO-END RESULTS]

Common Mistakes

Level-Up Tip

Keep one short department map document linking out to each agent's instructions and context pack. When the workflow changes, you update the map first and it tells you exactly which agent needs to change.

Action Checklist

Next Guide

20 — How to Stop Your AI Agent Making Things Up, where you'll add verification and review so a department you trust stays trustworthy.