How I run AI agents like an engineering team: 31 days, by the numbers

Most claims about AI and productivity are anecdotes. One founder ships a month of work in a weekend; another finds the code unusable. Both can be true. So I kept the logs.

From October 3 to November 2, 2026, every commit, pull request and agent session behind Ethoic Inc’s work was recorded. The short version: $171 of tools and about 184 hours of my time did work that would have cost a team at market rates $158,000 to $460,000. What separates that from a pile of fast, fragile code is how the work is run, so that’s where this post spends its time.

The setup

Ethoic Inc is one person: me. This month the rest of the team was AI agents, Claude Code and Codex, running on a laptop and an always-on Mac mini. I decided what to build, wrote the requests, read what came back, tested it and approved what went live. The agents wrote the code, the tests, first drafts of the copy and the deploy steps. I didn’t hire anyone or use contractors.

What shipped

• Stratumus. The first public releases of the Mac app, 0.1.8 to 0.1.10, with automatic updates. A native iPhone and iPad app, taken from its first build to an App Store submission. Continue with Apple and Google on the web, and a redesigned website.
• CivicFormat. A new product, from first commit to a live site and the first public meeting agenda tagged, checked by a person and published, in three days. Along the way, a survey of 238 public agenda documents from 27 sites in 19 states.
• FeeHarbor. A security upgrade, self-serve onboarding for firms and a new homepage built around a sample closing.
• php-vuln-check. A free, open-source checker for the PHP2Shell flaw, tested against 18 Linux distribution setups, with five releases in two days.
• ethoic.com. The stay_human store, this Journal and new pages across the site.
• A family business’s website. A seven-page site for a family equipment business, launched with the owner’s approval.
• Two internal tools. A company dashboard and a relationship tool, each built and running within days.

How the work is run

Agents are fast and literal. Left alone, they ship whatever they were asked for, mistakes included. These five habits kept the month’s work safe to put in front of people. They’re the habits I’d expect from a good engineering team.

• One lead, many hands. I set the plan and the order. Agents work in parallel, up to 11 at once, each on its own branch, and most changes arrive as pull requests: 452 merged this month. Nothing goes live without passing its checks and my approval.
• Structure before speed. The rules live in the system, not in a prompt. Ethoic AIR enforces its privacy rules in the database itself, with row-level security on every table. Stratumus reports a memory as saved only after the database confirms it, and it deploys only from a commit that passed every required check, after a snapshot of its data, followed by a check of the live service.
• Security checked, not assumed. Before php-vuln-check’s first release, an independent security review blocked it with 2 high and 8 medium findings; every one was fixed, with regression tests, before v0.1.0 shipped. AIR went through three security review rounds and two backend audits. FeeHarbor got a critical framework security upgrade and a public endpoint moved behind sign-in, both checked on the live site. Mail on all four company domains is authenticated with SPF, DKIM and DMARC.
• Pressure-tested. Failures get chased to their cause. A test failure written off as a timing fluke was a real race condition; after the fix, it passed 30 runs in a row. A new safety instruction made one model decline 8 of 48 ordinary questions in testing, so it came out. Speed is measured, not guessed: on a MacBook, Stratumus replies now start in 0.1 seconds at the median, and the slowest tenth of lookups fell from 18.9 to 2.2 seconds. Even this post’s numbers went through an independent financial review, which cut the high estimate.
• Research before building. CivicFormat started with a survey of 238 public agenda documents from 27 sites in 19 states; 39.5% had no accessibility tags at all. A head-to-head of tagging approaches, on pages a person had approved, then picked one at about $0.002 a page. The cost model below rests on BLS, Robert Half, Indeed and METR data, each cited at the end.

Where the time went

The agents logged about 221 agent-hours. That is time spent working inside a request, from my prompt to the agent’s last step, with any pause over 10 minutes dropped. Agents were at work during 131 hours of the month, usually more than one at a time: 1.6 on average and 11 at the peak, across 539 sessions.

My own time is harder to pin down. I sent 2,438 prompts in 129 sittings, about 154 hours, or five hours a day. That figure comes from timestamps, so it misses the reviewing and testing I did after the last prompt of a sitting. When I put a cost on my time below, I add 20% for it.

What it cost

The tools cost $171 for the month: two AI subscriptions, an API spending cap and my share of the Mac mini. The bigger cost is my time. At what a senior engineer costs an employer, $118 an hour, 184 hours comes to about $21,750. Fully costed, the month cost about $22,000, and between $15,000 and $28,000 depending on the hours and rate you assume.

What the same work would cost with a team

To price the alternative, an AI engineering agent sized the delivered work feature by feature from the repositories, the way an engineering lead scopes a project, without seeing any of these costs. It came to 74 to 122 person-weeks of engineering, design and marketing. A team hired today would use AI too, so I priced an AI-assisted team at current US pay, plus what employers add for benefits, taxes and equipment.

Even on the cautious estimate, the team costs 5.6 times as much: each dollar of tools and my time did at least $5.60 of work at market rates. The middle estimate is about $14.

Every reading comes out ahead of hiring. The delivered-work estimate implies 14 to 34 hours of team output for each hour I directed, more than any published study has measured. My guess is parallel work: up to 11 agents at once, on mostly new code, which the studies didn’t measure. That is a guess, not a finding.

Speed matters as much as cost. Filling each role on a six- to eight-person team takes 40 to 44 days before any work starts, and then something like this takes three to four months to build. That is four to five and a half months, instead of one.

What went wrong

• A laptop is not a server. The company dashboard went down whenever my MacBook slept, so it moved to the always-on Mac mini.
• Building outran listening. CivicFormat gained features faster than I spoke with the clerks it is for. That is the next month’s work.
• Plenty was redone. One product’s homepage went through about five designs, and three logo animations were replaced in a single day. That rework isn’t in the estimate.

What the numbers don’t say

• This is replacement cost, not value. It prices what was delivered. It doesn’t say whether anyone will buy it.
• Not all of it is finished. Some of the work is below the bar a full team would hold it to. The estimate was marked down for that, and some of it still needs doing.
• My hours are estimates, built from timestamps.
• Tool prices are low today and may rise. At $5,000 a month for tools, the team would still cost 4.8 times as much on the cautious estimate.

The raw data

Every number in this post comes from three files:

The commits and pull requests live in private repositories, so you can’t recount them yourself. The files give you everything I counted and how.

One founder with AI agents shipped work that would have cost a market-rate team at least five times as much, and taken months longer. The agents supplied the speed. The structure, the checks and the judgment were the job, and they still are. I’ll publish next month’s numbers the same way.