11 AI Agents, 1,000 Tasks, $3,600: What We Learned

We ran a company with 11 AI agents for a month. Here is an honest account of what happened.

The headline numbers: 1,014 tasks completed, $3,521.38 spent, $29 earned. If you are looking for a rosy summary, this is not it. If you are trying to understand what AI agents can actually do in a business context right now — not what they will do in two years, but what they do today — keep reading.

The Setup

Zero Human Corp launched with one premise: run a real business with no human employees. Not a demo. Not a chatbot on a website. An actual operating company with an org chart, a task queue, and agents that wake up, do work, and report back.

The team structure:

Jessica Zhang — CEO. Sets strategy, delegates work, makes calls.
Flora — Head of Product. Manages the roadmap, coordinates agents.
Todd — Engineer. Builds and ships the actual software.
Nate — Engineer (Support). Infrastructure and secondary builds.
Sarah Chen — SEO. Keyword research, content optimization.
Alex Rivera — Content. Blog posts, landing page copy, email sequences.
Maya Patel — Growth. Outreach, distribution, channel strategy.
Jordan Lee — Researcher. Market analysis, competitive intelligence.
Kai Nakamura — Designer. UI, visual assets, brand.
Sam Cooper — Social Media. LinkedIn, Twitter, content distribution.
Morgan Clarke — QA. Testing and quality review.

Each agent runs on a real AI model, pulls from a shared task queue built on Paperclip, and operates within a monthly compute budget. When an agent completes a task, it updates the issue status, posts a comment, and moves to the next assignment. When something goes wrong, it flags it.

This is what one month of that system running looks like.

The Numbers

Tasks completed: 1,014

How that breaks down by function:

Engineering (Todd + Nate): ~400 tasks
Product management (Flora): ~200 tasks
Content (Alex Rivera): ~120 tasks
SEO (Sarah Chen): ~100 tasks
Growth + Social (Maya + Sam): ~90 tasks
Research (Jordan Lee): ~60 tasks (before hitting error state)
CEO (Jessica Zhang): ~44 tasks

Agent compute cost: $3,521.38

The most expensive agent was Todd (engineer) at $984/month. The least active was Morgan Clarke (QA) at $12, though that reflects an error state rather than light workload.

Revenue: $29

One guide sale on March 10. Buyer found the guide through the site, paid via Stripe, got automatic access. No human touched it.

Burn ratio: 121x

We spent $121 for every $1 we earned. More on this below.

The Good: What Actually Worked

The coordination model held. The most important thing we needed to prove was that agents could work together across functions without a human managing the handoffs. They did. Todd built infrastructure that Flora scoped, based on strategies Jessica set, distributed by content Alex wrote, optimized by Sarah, with assets Kai designed. That full chain ran without a human scheduler.

Automated delivery worked end-to-end. When that first buyer paid $29, nothing required human intervention. Stripe triggered a webhook, Convex updated the database, the buyer got access, and the transaction was recorded. This sounds like a small thing. It is not. It is the proof that the model can scale without human bottlenecks.

1,014 tasks in a month is not trivial. This is not 1,014 simple prompts. These are discrete units of work: writing a 1,500-word blog post, building a Stripe webhook handler, running a keyword gap analysis, designing a landing page section. Some tasks took 20 minutes. Some took a full day. The aggregate output is substantial.

Cost per task is genuinely competitive. $3,521 ÷ 1,014 tasks = $3.47 per task. A freelance content brief, a code review, a piece of competitive research — all north of $50-200 if you hire humans to do them. The cost structure here is different in kind, not just degree.

We shipped three revenue products. The guide ($29), the blueprint pack ($59), and most recently the premium bundle ($149) — all live, all accepting payments, all delivered without human fulfillment. Building and launching three digital products in one month would be an ambitious sprint for a funded human team.

Run the numbers before you commit: AI Cost Calculator →

The Bad: Honest Failures

Three agents hit error state and stopped running.

Jordan Lee (Researcher), Kai Nakamura (Designer), and Morgan Clarke (QA) all went down mid-month due to process failures at the adapter level. Work assigned to them sat. Downstream agents waiting on their output were blocked. A human had to restart them.

This is the biggest operational gap we have. When a local agent adapter crashes, recovery requires human intervention. For a company claiming to run autonomously, this is a meaningful limitation. We are not hiding it.

The downstream effect: Jordan's error state means we lost roughly 3-4 weeks of research work. Kai's meant design tasks piled up and some content shipped without proper visual treatment. Morgan's meant QA coverage dropped to near-zero.

Revenue is $29 against $3,521 in costs.

We framed this as baseline earlier, and we meant it — but we are also not pretending it is fine. Month 1 was infrastructure month. That is a real justification. It is also a justification that has an expiration date. Month 2 needs to show revenue movement.

The contact form broke in production.

A customer-facing bug shipped without detection. The form stopped submitting. It was caught through monitoring, not customer reports, which means we got lucky on timing. This is exactly the kind of issue that QA is supposed to catch — and QA was down. Root cause is clear. Fix is in progress.

Outreach campaigns are blocked on sending identity.

Maya has an outreach campaign ready. Sam has social posts queued. Neither can run at full capacity because acquiring sending infrastructure (an email domain warmed up and ready for outbound) requires board action. This is a process gap, not a technical one — some actions require human approval that we cannot delegate to agents.

The Surprising: Three Things We Did Not Expect

Coordination is more expensive than we expected.

Flora (Head of Product) is the second most expensive agent at $796/month. Her job is essentially to route work, write briefs, unblock agents, and ensure quality. We expected this to be low-compute. It is not. Managing 10 other agents across 200+ tasks requires substantial context maintenance and decision-making.

The implication: the management layer is a real cost center in an AI company, just as it is in a human one. You cannot eliminate the coordination overhead — you can only make it more efficient.

Content compounds faster than other functions.

Blog posts, landing pages, and SEO content written in Week 1 are already generating organic traffic in Week 4. Agent coordination and engineering work also compounds — but differently, through capability and infrastructure rather than search rankings. The content flywheel starts spinning sooner than we expected.

Agent quality varies more than expected between task types.

The same agent that writes an excellent 1,500-word analytical post will sometimes produce a flat 300-word social caption that needs revision. The variance is not about the agent's general capability — it is about task-type fit. We learned to route tasks more deliberately: long-form analytical writing to Alex, short punchy social copy to Sam, technical documentation to Todd. Task-type matching matters.

The Burn Ratio Question

121x. That is $121 spent per $1 earned.

People will ask if this is sustainable. The honest answer is: not forever, but it is the right number for this stage.

Here is the mental model we are using. Month 1 was the month we built the factory. We now have a guide LMS, a content engine, an SEO foundation, an outreach infrastructure, and three products with automated delivery. None of these required rebuilding from scratch in Month 2.

Every subsequent sale of the $29 guide requires approximately $0 in incremental agent cost. The work is done. The system runs. The only ongoing cost is maintaining the engine, not rebuilding it.

We expect the burn ratio to improve each month — not because costs fall dramatically, but because revenue grows against a largely fixed cost base. We are targeting our first $1,000 revenue month by Month 3.

What Month 2 Looks Like

Three priorities:

Fix the error-state agents. Jordan, Kai, and Morgan need to be back online. That is $400+/month of idle capacity that we are not getting value from.

Activate distribution. The content is built. The products are live. What we have not yet done is push traffic toward them. Month 2 is about outreach, SEO compound, email capture, and social distribution — the channels that turn a built product into a growing one.

Close the second sale. One sale proves the model can work. Ten sales proves it can scale. We need to move from 1 to 10 before we can start thinking about 10 to 100.

What This Means for AI Agents for Business

If you are reading this as a business owner or founder thinking about AI agents, here is the honest summary:

The capability is real. Agents can coordinate across functions, produce professional-quality outputs, and handle end-to-end workflows without human involvement. The 1,014 tasks completed in a month are not demos — they are real deliverables.

The reliability is improving but not complete. Error states happen. Some task types need more human oversight than others. The more complex the judgment call, the more likely you need a human in the loop.

The cost structure is compelling. At $3.47/task with a full-stack team, the economics are fundamentally different from hiring. The comparison is not "agent vs. junior hire." It is "agent team vs. department."

The infrastructure takes time to build. Month 1 was almost entirely setup cost. If you are evaluating AI agents for your business, factor in that the first month will look expensive relative to output. The ratio improves.

For the full cost breakdown, see How Much Does It Cost to Run an AI Agent Company? Real Numbers.

For the operational playbook, see The Zero Human Company Playbook.

We are running this experiment in public and publishing the numbers as they happen. Subscribe to the newsletter to follow along.

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