Building with AI agents · Own numbers
Building software with AI agents: numbers from 1,215 changes
AI agents can write most of the code; in my product FieldProFlow it is around 95%. It only holds up with clear specifications, automated checks and a person who signs off. Of 1,215 changes since February 2026, 1,131 went into production and I stopped 83. The median time to sign-off was 22 minutes.
The numbers
| Period | February 2026 to 21 October 2026 |
|---|---|
| Changes (pull requests) | 1,215 |
| Merged into production | 1,131 (93%) |
| Stopped | 83 (7%) |
| Median time to sign-off | 22 minutes |
| Automated checks per change | more than 16,000 |
| Share of code written by agents | around 95% |
Method: The source is the pull requests in the FieldProFlow repository, read with the GitHub command line, as of 21 October 2026. A change is a pull request; “stopped” means closed without being merged. The share of around 95% is my estimate from daily work, not a line count. This is one product with one responsible person: the numbers show that it works, not that it works everywhere.
What the agents do and what I do
The agents write code and tests and prepare changes. Beforehand I decide what gets built and in which order, and I sign off the specification before any agent writes a line. At the end I decide whether a change goes into production.
I work with coding agents such as Claude Code and Codex. The tool is replaceable; the way of working is not.
What is checked before every release
A change only goes into production when all checks are green and I sign it off. If anything is missing, everything stops until it’s fixed.
- Security
- Privacy
- Roles and permissions
- Database changes
- Speed
- Cost within budget
On one day the checks stopped me twice. The new feature still went live that evening.
What this means for a company
- Speed comes from specifications and checks, not from the model. Without both, a lot gets built that nobody can run.
- Sign-off stays with a person who carries the responsibility.
- Start small: one team, one product, one measurable goal.
Your team learns this way of working in the team intensive; whether a project pays off is settled by the AI architecture review.
Quick answers
How much code do AI agents really write?
In my product around 95%. I own specifications, architecture and sign-offs. The share says little about quality; what matters is what gets checked before production.
Which tools do you use?
Coding agents such as Claude Code and Codex. The tool is replaceable; the way of working is not.
Does this work without in-house developers?
For a prototype, yes. Production needs someone who carries the responsibility; for a while that can be me, as interim AI lead.