★How I build
I don't write code by hand. I direct AI agents, and I verify everything they do.
Typing code line by line is the slow part of software. My work is the part that decides whether software is right: specifying the problem, splitting it into tasks, delegating each one to an agent, reviewing every change and gating it with tests. Everything on this page, and all of Deliservy, was built that way.
The cost of writing code by hand
- Slow. The standard cost model puts a full team at about 28 months for this much code.
- Expensive. 21 to 37 salaried engineers, $12.0M–$39.3M before management, design and QA.
- Tests get skipped. When every test costs a person's afternoon, coverage is the first thing cut.
- Knowledge stays in heads. Decisions live with whoever typed the code, and leave with them.
- One task at a time. A developer works on one thing; changing direction is costly.
What directing agents changes
- Speed. ~185k lines across the app, backend, panels and website in about 6.5 months, with one engineer.
- Cost. About $245k–$380k (6.5 months of a Staff/Principal engineer's pay, plus ~$1,600 of AI subscriptions) instead of a department.
- Tests by default. 5,343 automated test cases, written alongside the features, not after.
- Parallel work. Up to six agents on separate branches at the same time.
- Written down. Rules, design notes and decisions live in the repository, where every agent reads them.
See the full numbers: cost, time, team size and the AI bill
| Writing the same ~185k lines | People | Time | Cost |
|---|
| By hand, lean team (COCOMO) | ~21 engineers | ~28 months | $12.0M–$21.8M |
| By hand, typical team (COCOMO) | ~37 engineers | ~28 months | $21.6M–$39.3M |
| Directing AI agents (what actually happened) | 1 Staff/Principal engineer | ~6.5 months | ~$245k–$380k |
| Saved | 20–36 people | ~21 months | $11.6M–$39.0M (97–99%) |
| Codebase | Lines written with agents | Test cases | By hand, lean estimate |
|---|
| Mobile app (iOS and Android) | ~80k (98k total) | 1,994 | ~$5.0M–$9.0M |
| Backend, bank and AI integrations, scripts | ~62k (69k total) | 2,996 | ~$3.9M–$6.9M |
| Admin service, admin panel and parcel panel | ~40k (41k total) | 353 | ~$2.4M–$4.4M |
| Website | ~2k (4k total) | – | ~$0.2M–$0.3M |
COCOMO (basic) is a model, not a quote. It is applied to the ~185k lines of production code written directing AI agents (blank lines, comments, generated code and third-party libraries excluded) and priced at the pay of a senior software engineer at a U.S. tech company ($250k–$350k a year in total compensation), plus ~30% overhead at the high end. Per-codebase figures are estimated separately and add up to slightly less than the whole. "What actually happened" is priced at the pay of a Staff to Principal engineer in the U.S. ($450k–$700k a year in total compensation: the level of someone who designs, builds and runs whole systems alone) for about six and a half months, about $244k–$379k, plus about $1,600 of AI subscriptions. The traditional teams are priced at senior pay, not at Staff/Principal pay, so the comparison favors them. AI-written code and UI markup can be wordier than hand-written code, so a line-based model may overstate the manual effort; read these as an order of magnitude.
Alisa, the AI assistant, measured the same way
| By hand (rough COCOMO estimate) | 2–4 engineers · ~8 months · $0.2M–$0.4M |
| Directing agents (actual) | 1 Staff/Principal engineer · ~7 weeks of focused work · ~$30k–$47k |
Alisa is about 10,900 lines (core, channels and voice, comments included). The estimate assumes the same code-to-comment ratio as the whole backend.
How I keep the AI bill small
- Flat subscription for building. All the coding agents ran on one Claude Max plan at $200 a month: about $1,600 for the whole build, around 2% of what it actually cost.
- The right model for each task. Faster models do routine work; the strongest ones are saved for hard problems and reviews.
- Hard limits in production. The public version of Alisa has per-IP limits and a daily message cap, so a spike can't run up the bill.
- Fewer wasted runs. Clear specs and shared rules mean agents get it right sooner, and tests catch mistakes before they cost a second attempt.
6agent sessions working in parallel at peak
13agents at once at peak: 1 session directing 12 sub-agents
186agent sessions logged since 13 Aug 2026 (61 main, 125 sub-agents)
3layers of verification: tests, reviewer agents and me
Measured from the session logs kept on my computer, which go back to 13 Aug 2026; earlier months are not counted.
1 · Machines check
The agent that writes a change runs the type checker and the test suite before it reports back. Nothing is marked done while a test fails.
2 · Agents check each other
For critical work, separate reviewer agents that didn't write the code try to break it and must reproduce every bug they report. The delivery engine on this page went through three such rounds.
3 · A human checks everything
I write the specs, read the diffs, test on real phones and in the emulator, and approve every deploy. Anything that moves money or messages a customer needs my explicit yes.
Advanced skills
Prompt engineeringAgent delegationMulti-agent orchestrationSpecs and acceptance criteriaCode review of agent outputTest design and verification gatesIndependent agent reviewsMutation and fuzz testingModel selection by taskToken and cost controlDeployment and production safetyAgent memory and shared rules
What I can do for a company
- Cut engineering costs by moving teams from hand coding to directed agents, with the review and test gates that keep quality up.
- Train engineers who are not yet clear on how to prompt or delegate: how to write a task an agent can't misread, and how to catch it when it does.
- Set up the structure: shared rules, memory, branch strategy and review rounds, so agents work in parallel without breaking production.
Where this is going
My bet: within about three years, typing code will matter far less than knowing how to structure the work. That means clear specs, good delegation, strict verification and safe deployment for software, apps, servers and AI systems. That is the skill I practice every day, and I'm already working that way in production.
