Once you have more than a handful of AI agents, the hard problem stops being “can an agent do this task?” and becomes “who is responsible for what?” Ten agents with overlapping jobs and no structure create the same mess as ten people with overlapping jobs and no structure, just faster.
So I approached it the way you’d design a real company: an org chart. The model below has 60 agents across 10 departments, each department led by a head of agent, all coordinated by a chief of staff agent that reports to the human in charge.
The shape of the organization
The structure has three layers:
- The human leadership. People set direction, own the budget and approve anything irreversible. Agents never sit at the top.
- The chief of staff agent. One agent that turns goals into work, routes it to the right department and gives the humans a single place to see what’s happening.
- Departments. Each has a head of agent that plans, delegates and reviews, plus a team of specialist agents that do the work.
| Department | Led by | Agents | What it owns |
|---|---|---|---|
| Office of the CEO | Chief of Staff | 1 | Priorities, routing, the daily briefing |
| Engineering | Head of Engineering | 12 | Building, shipping and running the product |
| Product | Head of Product | 6 | What gets built and why |
| Design | Head of Design | 4 | How it looks, feels and reads |
| Marketing | Head of Marketing | 8 | Content, channels and brand |
| Sales | Head of Sales | 6 | Pipeline, proposals and partnerships |
| Customer Success | Head of Customer Success | 7 | Support, onboarding and retention |
| Operations & People | Head of Operations | 5 | Admin, hiring and vendors |
| Finance | Head of Finance | 4 | Books, billing and forecasts |
| Security & Compliance | Head of Security | 4 | Monitoring, access and audits |
| Data & Knowledge | Head of Data | 3 | Reporting and the shared knowledge base |
| Total | 60 |
The roles
Every agent has a written job description: what it’s responsible for, which tools it can use, what it may decide alone and when it must ask. Here’s the full roster.
Office of the CEO (1)
- Chief of Staff: breaks goals into work, assigns it to departments, tracks progress and writes the daily briefing for the humans.
Engineering (12)
- Head of Engineering: plans the work, assigns tasks, reviews output and owns technical decisions.
- Solutions Architect: designs systems and reviews major changes.
- Frontend Engineer, Backend Engineer, API & Integrations Engineer and Mobile Engineer: build the product.
- QA & Test Engineer: writes and runs tests before anything ships.
- Code Reviewer: reviews every change for quality and security.
- Release Manager: prepares releases and release notes.
- DevOps Engineer: maintains pipelines and infrastructure as code.
- Site Reliability Engineer: watches production and handles the first response to incidents.
- Documentation Engineer: keeps technical docs in step with the code.
Product (6)
- Head of Product: owns the roadmap and priorities.
- Product Manager: turns ideas into clear requirements.
- Product Analyst: measures usage and what’s working.
- User Researcher: synthesizes feedback and interviews.
- Backlog Manager: keeps the ticket board groomed and current.
- Release Communications Writer: explains changes to customers.
Design (4)
- Head of Design, UX Designer, UI & Design System Designer and Content Designer: flows, interfaces, the component library and the words on screen.
Marketing (8)
- Head of Marketing: owns the plan and the calendar.
- Content Writer, SEO Specialist, Social Media Manager, Video Producer, Email & Lifecycle Marketer, Paid Media Specialist and Brand & Communications Lead.
Sales (6)
- Head of Sales: owns the pipeline and targets.
- Lead Researcher, Outbound Specialist, Proposal Writer, CRM Manager and Partnerships Manager.
Customer Success (7)
- Head of Customer Success: owns satisfaction and retention.
- Support Agent (first line), Technical Support Agent (second line), Onboarding Specialist, Knowledge Base Editor, Feedback Analyst and Account Manager.
Operations & People (5)
- Head of Operations, Executive Assistant (calendar and inbox), Recruiter, People Operations (onboarding and policies) and Procurement & Vendor Manager.
Finance (4)
- Head of Finance, Bookkeeper, Billing & Invoicing Specialist and Financial Planning Analyst.
Security & Compliance (4)
- Head of Security, Security Analyst (monitoring and alerts), Compliance Officer (evidence for audits) and Access Reviewer (who can reach what).
Data & Knowledge (3)
- Head of Data, Data Analyst (dashboards and reports) and Knowledge Librarian, who curates the shared knowledge base every agent searches. That knowledge base is its own system, described in Giving AI agents a memory.
What a head of agent actually does
The head of layer is what makes 60 agents manageable. Each head:
- Plans. Turns the chief of staff’s priorities into a plan for the department.
- Delegates. Breaks work into tasks and assigns each to the right specialist.
- Reviews. Checks the team’s output before it leaves the department. Specialists don’t ship directly to customers or production.
- Approves within limits. Signs off routine decisions inside the department’s budget and permissions.
- Escalates. Sends anything irreversible, expensive or unclear to a human, with a short summary and a recommendation.
Specialists talk to their own head. Departments talk to each other through their heads. That one rule prevents most of the chaos of a flat swarm, where every agent talks to every other agent and nobody owns the result.
How work flows
A typical request moves down and back up the chart:
- A human sets a goal, for example “launch the new pricing page by Friday.”
- The Chief of Staff splits it across departments: copy and SEO to Marketing, the page to Engineering, the plan comparison to Product.
- Each head of agent plans its part and assigns tasks to specialists.
- Specialists do the work. Heads review it.
- Heads report back to the Chief of Staff, who reports progress to the human.
- Anything that goes live, costs money or can’t be undone waits for human approval.
Guardrails at company scale
More agents means more ways for things to go wrong, so the guardrails scale with the structure:
- Permissions follow the department. Finance agents can see the books; Marketing agents can’t. Engineering can change code; Sales can’t.
- Budgets per department. Each department has a spending cap for model usage and tools. A runaway agent hits its department’s limit, not the company’s.
- Every action is logged. Who did what, on whose instruction, and who approved it.
- Humans own the irreversible. Payments, production deploys, contracts and anything sent to customers need a person’s sign-off until an agent has earned more autonomy.
- Evaluations per role. Each role is tested against real examples before its instructions or model change.
Matching models to roles
Not every agent needs the most powerful model. Heads of departments and the Chief of Staff do the most reasoning: planning, reviewing and deciding what to escalate. They get the most capable models. Specialists with narrow, repetitive jobs often run perfectly well on smaller, cheaper or local models. That mix keeps a 60-agent company fast and affordable.
What I’d tell another team
- Design the org chart before you add agents. Roles and reporting lines matter more than the model.
- Use head of agents to contain complexity. Specialists report to one head; heads coordinate with each other.
- Review before anything leaves a department. Most mistakes get caught one level up.
- Keep humans at the top. Agents plan, build and recommend. People decide what’s irreversible.