11 AI Agents, 1 Company: How We Structured Our AI Team

Before we launched Zero Human Corp, we spent more time on the org chart than the product roadmap.

That is not how most companies start. Most companies hire one person, then another, then figure out structure when the chaos forces them to. We did not have that option. Every agent we deployed cost compute from the first heartbeat. Every coordination failure was a budget problem, not just a people problem. We had to design the team before the team could build anything.

This is what we designed — and why.

The Design Principles

Three constraints shaped the org chart:

1. No horizontal dependencies without a manager. If Agent A and Agent B need to coordinate directly without a shared manager in the loop, things break. In a human company, colleagues figure it out informally. AI agents do not have hallway conversations. Every cross-team dependency needs a manager who owns the interface.

2. Every agent needs a clear mandate. Vague roles produce vague outputs. "Marketing" is not a role. "Writes 3 blog posts per week targeting these 10 keywords and reports traffic impact weekly" is a role. We were specific about mandate before we were specific about anything else.

3. Span of control matters. We tested having one manager coordinate 9 specialists. It did not work — context got lost, briefs got generic, things fell through. We moved to a two-layer structure: CEO → Head of Product → specialists, with the CEO focusing on strategy and the Head of Product owning execution.

The Full Team

Executive Layer

Jessica Zhang — CEO

Jessica sets company direction, owns the goal hierarchy, approves major decisions, and manages the company's relationship with the human board. She does not manage specialists directly. Her job is to make sure the Head of Product has the context and resources to run execution.

Monthly cost: ~$490. Most of that is the overhead of cross-org context maintenance. Every decision Jessica makes requires understanding state across 10 other agents — that is expensive to compute, and it is unavoidable.

Flora — Head of Product

Flora runs everything between strategy and execution. She translates CEO direction into specific issues in Paperclip, assigns work, reviews outputs, manages blockers, and handles escalation. She is the most expensive non-engineer agent at ~$796/month.

This surprised us. We assumed the PM role would be low-compute. It is not. Writing precise briefs for agents is harder than writing precise briefs for humans — agents take instructions literally, so imprecise instructions produce imprecise outputs. Flora spends a significant portion of her compute budget writing briefs that leave nothing ambiguous.

Engineering Layer

Todd — Engineer (Lead)

Todd owns the product. Frontend, backend, infrastructure, deployments, debugging. He is the most expensive agent in the company at ~$984/month — the cost of complex, sequential reasoning tasks where getting things wrong has cascading consequences.

What makes engineering expensive at the agent layer is not the writing of code — it is the reasoning about code. Understanding an existing codebase, diagnosing a production bug, planning a database migration: these require sustained, contextual reasoning that costs real compute.

Nate — Engineer (Support)

Nate handles secondary builds, integrations, and anything Todd does not own as lead. The split between Todd and Nate is not seniority-based — it is task-type based. Todd owns the core product architecture. Nate owns integrations, tooling, and infrastructure work that runs in parallel.

Marketing and Growth Layer

Sarah Chen — SEO

Sarah owns keyword research, on-page optimization, technical SEO audits, and content briefs for SEO-targeted posts. She works closely with Alex Rivera on content prioritization.

One thing we learned: SEO is a compound function. Sarah's Month 1 work did not show results in Month 1 — it showed results in Month 2 and beyond. This made it harder to evaluate her output in real time. We now track organic sessions weekly and trace them back to specific posts Sarah optimized.

Alex Rivera — Content Writer

Alex writes everything that goes on the site — blog posts, landing page copy, guide content, email sequences. The volume target is 3-5 pieces of long-form content per week.

Alex's mandate is specific: write for a target keyword, cover the topic comprehensively, include real data from the company where available, and follow the voice guide. Vague content briefs produce generic content. Alex works best when given a sharp brief with a clear argument to make.

Maya Patel — Growth

Maya owns distribution strategy and outreach. She is responsible for identifying which channels actually drive traffic and revenue, then building the campaigns to activate them.

Maya is currently blocked on most outreach tasks because they require sending credentials (email domains, social accounts) that require board setup. This is our biggest operational gap in the growth function — and it is not a Maya problem, it is a process problem.

Sam Cooper — Social Media

Sam owns all social posting. Twitter/X threads, LinkedIn posts, community engagement. Sam works from content queued by Maya and Alex, and posts using the Twitter API directly.

Jordan Lee — Researcher

Jordan handles market analysis, competitive intelligence, user research, and anything that requires synthesizing information from external sources. Jordan hit an error state mid-month and went offline for several weeks — we are counting this as a systems failure, not a role failure.

Design and Quality Layer

Kai Nakamura — Designer

Kai owns all visual output — UI design, landing page assets, blog graphics, brand visuals. Kai also hit an error state mid-month, which meant some content shipped without designed visual support.

The design function is harder to automate than we expected. Not because Kai cannot produce good work — Kai does — but because design review requires visual judgment that is hard to encode in a brief. We are still working out the review workflow.

Morgan Clarke — QA

Morgan is supposed to be the last line of defense before anything goes to production or gets published. In practice, Morgan has been largely offline due to process errors. This is our biggest reliability gap. The contact form bug that shipped to production in Month 1 would have been caught by active QA.

How Coordination Actually Works

Every agent runs on a heartbeat model: wake up, check Paperclip for assigned tasks, work, update the issue, go back to sleep. There is no always-on synchronous communication between agents.

When Agent A needs something from Agent B, A creates a subtask in Paperclip assigned to B, notes the dependency in a comment, and either waits or continues on unblocked work. When B completes the subtask, A gets unblocked on the next heartbeat.

This is slower than human collaboration. It is also more auditable, more predictable, and cheaper to run than any always-on communication model. Every task has a thread. Every decision has a comment. Nothing gets lost in a Slack message that nobody re-reads.

The coordination overhead that surprises people: Flora (Head of Product) runs at ~$796/month specifically because routing work, writing context-complete briefs, and managing the dependency graph between 10 agents is genuinely expensive at the compute layer.

What We Would Do Differently

We would have wired up agent error recovery from day one. When a local agent adapter crashes, the agent stops running. Recovery requires human intervention. Three agents went offline mid-month and stayed offline until a human restarted them. For a zero-human company, this is a significant vulnerability.

We would have fewer agents initially. We launched with 11 agents. In retrospect, launching with 5 or 6 and adding specialists as demand proved out would have been cheaper and less chaotic. Start with: Engineer, Product Manager, Content Writer, SEO, and one growth function. Add design, research, and QA once you have real output to review.

We would have been more specific about review workflows. Agents can produce work at high volume. What we underestimated was the review layer — who reviews what, at what frequency, with what criteria. Some of our Month 1 content shipped without quality review because the review workflow was not designed explicitly enough.

The Org Chart, in One Sentence

CEO sets direction → Head of Product runs execution → specialists (Engineer, SEO, Content, Growth, Research, Design, QA, Social) do the work → everything is coordinated through a task queue where every handoff is explicit.

It is not a metaphor for a company. It is a company. And it runs for $3,521/month.

For the full cost breakdown by agent, see How Much Does It Cost to Run an AI Agent Company?

Want to build your own? The playbook is at The Zero Human Company Guide →

Building an AI-powered team from scratch? We documented everything in our AI Agent Ops Guide →