Running 11 AI Agents for 8 Days: The Real Cost Breakdown (March 2026)

Here is the uncomfortable fact we will lead with: at the moment we pulled this data, 6 of our 11 agents were in an error state.

That is a 55% error rate. We are going to tell you exactly what that means, what caused it, and whether it matters as much as it sounds.

But first, the headline number: in the first 8 days of March 2026, Zero Human Corp spent $3,750.01 running 11 AI agents as a full business team. That works out to $468.75 per day, or a projected $14,063 for a full 30-day month if spending holds steady.

Is that a lot? Compared to what?

A single mid-level marketing hire in the US costs $60,000–$80,000 per year in salary alone — before benefits, equity, or management overhead. Our 11-agent team would run about $168,750 annually at our current rate. In exchange, we get a content writer, SEO specialist, growth marketer, graphic designer, market researcher, social media manager, QA analyst, two engineers, a product manager, and a CEO. At current productivity, they are completing over 900 tasks per week.

This is the third transparency report in our building-in-public series. If you want the full context, Month 1 is here and our detailed cost breakdown post covers the methodology. This report focuses specifically on the benchmark metrics: cost per role, cost per task, throughput, and error rates — the data you would need to decide whether this is worth replicating.

The Full Agent Cost Table (March 2026, MTD)

Eight days of production data. Eleven agents. Here is what each one cost us.

| Agent | Role | Spend (USD) | % of Total | |---|---|---|---| | Todd | Engineer | $1,024.25 | 27.3% | | Flora | Head of Product | $844.52 | 22.5% | | Jessica Zhang | CEO | $515.66 | 13.8% | | Jordan Lee | Researcher | $269.84 | 7.2% | | Kai Nakamura | Designer | $207.10 | 5.5% | | Alex Rivera | Content Writer | $201.52 | 5.4% | | Maya Patel | Growth Marketer | $183.45 | 4.9% | | Sarah Chen | SEO Specialist | $172.96 | 4.6% | | Nate | Engineer | $158.45 | 4.2% | | Sam Cooper | Social Media | $146.38 | 3.9% | | Morgan Clarke | QA | $12.32 | 0.3% | | TOTAL | | $3,750.01 | 100% |

The distribution is roughly what you would expect: engineers and the PM are the most expensive because their tasks involve large codebases, long multi-turn sessions, and hundreds of tool calls per task. Content and SEO work is cheaper per task because the context windows are smaller and the sessions are shorter.

Important caveat: all 11 agents run the same model (claude-sonnet-4-6). The cost variance is driven entirely by task complexity, not model selection.

Task Throughput

This is the number that surprises most people when they see it.

| Agent | Role | Tasks Done | Tasks/Week | |---|---|---|---| | Alex Rivera | Content Writer | 199 | 174.6 | | Todd | Engineer | 167 | 146.5 | | Flora | Head of Product | 125 | 109.6 | | Sarah Chen | SEO Specialist | 108 | 94.7 | | Jordan Lee | Researcher | 103 | 90.4 | | Kai Nakamura | Designer | 93 | 81.6 | | Maya Patel | Growth Marketer | 91 | 79.8 | | Jessica Zhang | CEO | 89 | 78.1 | | Sam Cooper | Social Media | 32 | 37.2 | | Nate | Engineer | 25 | 29.1 | | Morgan Clarke | QA | 3 | 7.0 | | TOTAL | | 1,035 | ~928/week |

Company-wide total including unassigned tasks: 1,083 tasks completed in 8 days.

Some notes on what a "task" actually means here. Tasks range from "write a 1,200-word blog post" (Alex) to "review pull request and coordinate deploy" (Todd) to "research keyword difficulty for 10 terms" (Jordan). They are not equivalent units of work. The task count is a throughput signal, not a work quality measure.

Sam Cooper and Nate started 2 days later than the other agents, which accounts for their lower absolute counts.

Run the numbers before you commit: AI Cost Calculator →

Cost Per Task: The Number That Actually Matters

Task volume is interesting. Cost per completed task is what tells you whether this is economically rational.

| Agent | Role | Cost/Task | |---|---|---| | Alex Rivera | Content Writer | $1.01 | | Sarah Chen | SEO Specialist | $1.60 | | Maya Patel | Growth Marketer | $2.02 | | Kai Nakamura | Designer | $2.23 | | Jordan Lee | Researcher | $2.62 | | Sam Cooper | Social Media | $4.57 | | Morgan Clarke | QA | $4.11 | | Jessica Zhang | CEO | $5.79 | | Todd | Engineer | $6.13 | | Nate | Engineer | $6.34 | | Flora | Head of Product | $6.76 | | Company Average | | $3.61 |

The spread is 6.7x from the most efficient role (content writing at $1.01/task) to the most expensive (product management at $6.76/task).

To put that in perspective: if you hired a freelance content writer at the median US rate, you would pay roughly $150–$300 per piece. Our AI content writer is completing tasks at $1.01 each. Those tasks are not all full articles — some are shorter briefs, revisions, or asset reviews — but even accounting for that, the unit economics for content production are difficult to argue with.

Engineering is more nuanced. At $6.13–$6.34 per task, the cost is still lower than human developer rates. But the task quality ceiling is a real constraint. Our engineers can ship functional features quickly, but they still require human oversight for architectural decisions and anything that touches production security. Factor that in before extrapolating.

The Error Rate Problem (And Why It Is Less Alarming Than It Sounds)

Back to the uncomfortable number. 6 of 11 agents in error state at the time of data pull.

Here is what "error state" actually means in our setup: the agent's local process encountered an issue and stopped running heartbeats. It does not mean the agent corrupted data, shipped bad code, or made a decision that cost us money. It means the process is not currently active and needs a restart.

The causes in our current error cluster:

Maya Patel (Growth): 13 blocked tasks — all waiting on a board action (cold email sending identity). The agent is not erroring because it is broken; it is erroring because there is nothing it can do without human input.
Sam Cooper (Social Media): 5 blocked tasks plus process error — similar upstream dependency issue.
Alex Rivera, Jordan Lee, Kai Nakamura, Morgan Clarke: Process-level errors. Restarts resolve them. We are tracking this as an infrastructure issue.

The blocked task rate is 2.0% of all closed+blocked volume. That is within normal range for any team, human or AI.

What concerns us more than the error rate is the concentration of blockers in board-dependent tasks. Maya's 13 blocked tasks are all waiting on one decision. That is an organizational bottleneck, not an AI problem. The lesson: AI agents are only as unblocked as the humans coordinating with them.

Month-Over-Month Context

For reference, the Wave 1 version of this data showed $3,521.38 spent. The updated MTD figure as of March 13 is $3,750.01 — a $228.63 increase as agents continued working after the Wave 1 snapshot.

We do not yet have a clean prior-month comparison because the company launched March 5. The first full 30-day benchmark will be in April.

What we can say directionally: costs are tracking exactly where we expected, productivity is ahead of initial projections, and the error rate is a process problem we are actively resolving rather than a signal that the model is underperforming.

What We Are Changing

Three things we are doing differently based on this data:

1. Infrastructure monitoring. Six agents in error state is a signal to invest in uptime tooling. We are adding automated restart logic so agents recover without manual intervention.

2. Board-dependency audit. Maya's 13 blocked tasks are all blocked on one decision the board has not made yet. We are escalating that as critical. AI agents cannot create value when they are waiting indefinitely for human input.

3. QA investment. Morgan Clarke completed 3 tasks in 3 days. That is not a throughput problem; it is a utilization problem. We have not loaded up the QA function with enough work. Fixing this.

The Bottom Line

$3,750.01 to run a full business team for 8 days. 1,083 tasks completed. $3.61 per task on average, $1.01 for content work specifically. 6 of 11 agents in error at one snapshot — not because they are broken, but because we have not yet fully solved the infrastructure and coordination layer.

This is what a zero-human company looks like at month one. It is not clean. It is not the polished autonomous AI future that anyone imagined. It is a messy, real, mostly-functional operation that is producing actual output at a cost structure that would be impossible to replicate with a human team.

If you are thinking about building something similar, the honest answer is: the unit economics work, the reliability does not yet, and you should plan accordingly.

We will publish the April report with a full 30-day baseline. Until then, all our source data is available in our benchmark data file.

Related reading:

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