Selected workInternal tooling · AI agents
Same quality.A third of the cost.
My software factory turns a plain-English request into a reviewed, ready-to-merge change on a client website. In October 2026 I rebuilt it, measured every step, and cut its cost by 63% without losing a single graded job.
The results.
Ten realistic website requests, each run twice, before and after the rebuild. Every job was graded on what it actually left behind.
- Jobs finished correctly
- 20 of 20 before, 20 of 20 after. Same quality.
- Cost for the 20 jobs
- $31.25 before, $11.69 after: 63% lower.
- AI model calls
- 1,525 before, 731 after: 52% fewer.
- Time to finish
- 90 minutes before, 34 minutes after: 62% faster.
- Blocked or refused actions
- 38 before, 18 after: 53% fewer.
- Smallest jobs
- A one-line copy edit went from about $1.60 to about $0.47.
Costs are API-equivalent figures from the runs. The work itself ran on a flat subscription, so read them as relative, not as an invoice.
The problem.
The factory used to pass each request through seven separate AI agents: a spec writer, a builder, five reviewers and an approver. Every hand-off repeated context, and every agent paid for it.
I rebuilt it as one agent that plans and builds, three focused reviewers, and a plain script for the final go or no-go.
- Claude Code
- Agent design
- Evals
- Node.js
Measured first.
Before changing anything, I wrote 10 realistic site requests: copy edits, a new page, a phone number change, a contact form, an added package, and a cleanup with 18 design errors.
I ran each one twice and graded the result, never which agents ran. That gave me the baseline every change had to beat.
One change at a time.
- 1. The planner
- Merged into the main agent.
- 2. The builder
- Merged into the main agent.
- 3. Design, UX and SEO review
- Three reviewers became one reviewer with three checklists.
- 4. The approver
- The AI approver became a script with tests. It checks scope, protected files, new phone numbers or prices, the build and the done check.
After each change I re-ran all 20 jobs, compared them to the baseline, and decided to keep it, fix it or roll it back.
Caught my own mistake.
One fix made things worse: 12 of 20. The eval caught it the same day. I traced the cause in the logs, and the corrected version went back to 20 of 20 at the lowest cost of any run.
Rules in code, not prompts.
“Never touch the deploy files”, “a new phone number or price goes to the owner” and “the build must pass” are now checked by a script that can’t be talked out of them.
Max and the crew are AI agents. Every change they make is still reviewed and signed off by me before it reaches a client site.
Why it matters to you.
The same discipline goes into your site: measure before changing, change one thing at a time, and keep a person in charge of anything risky like prices, contact details or new dependencies.
Tell me about your project