Skip to main content
Software development / AI

How we build software: AI agents under fixed rules

We use AI agents in development, but under rules a person sets. Agents build, review and test every ticket in separate roles, and nothing is merged unless all required checks are green. The process is run from a development server of our own, from a phone as well.

← Back to all case studies
One person
sets the rules; agents can only propose rule changes
4 required checks
green on the exact commit or no merge, admins included
No write access
for agents to production systems and their data
30 minutes
of watching after every release: health routes and error logs

Starting point

AI agents can now take on a large part of the work on a ticket: writing code, adding tests, reproducing a bug. In many teams that means, in practice, an agent with a terminal, write access to the repository and, if need be, access to production. For our clients that is not an option. Anyone who builds software for consultancies in compliance is asked first who can reach production, who checked a change before it was released, and how that can be shown.

We wanted the throughput agents make possible without leaving those questions open. That took two things: a process whose limits are written down and enforced technically, rather than left to a model's judgement, and a working environment in which a person can see that process at any time and step in, even when away from the desk.

Solution

The process has a single rulebook. It states what the agents may do without asking, what they never do whatever a ticket says, and when they stop and ask. No agent writes to a production system or any other deployed instance, touches its credentials, pushes straight to a protected branch, sends real e-mail or calls paid third-party services. Only a person can change the rulebook. A separate supervisor session watches the process, measures waiting times and load and proposes rule changes, but never builds anything itself; a proposal applies only once the person explicitly confirms it.

Every ticket passes through separate roles. A coordinator picks it by fixed rules and writes the brief. A build agent works in a git worktree of its own and in a sandbox without network access, where credentials and environment files are hidden. Tests, lint, type checks and an architecture check run outside the sandbox, and only when they are green is anything committed. A second AI model then reviews the diff in at most two rounds, more thoroughly for migrations, permissions and tenant boundaries. Which model takes which role is the person's decision; the process never switches it on its own.

Testing never happens on a shared or deployed system, but on dedicated QA environments: copies of the development environment with their own database, their own mail catcher and no public access. A QA agent walks through the change in a headless browser, the way a person would. Every expected value in the test brief names its source in the code, such as the translation key of a label or the route of a path; a value without a source does not go in. Test data is recorded by its identifiers and deleted by exactly those, never by a name pattern.

Merging happens only through pull requests, and only when all four required checks are green on that exact commit; the branch rules apply to administrators as well. Releases run on working days within a fixed time window, one at a time and at least two hours apart. After every release the process watches the health routes for 30 minutes and compares the error logs with the half hour before. If a route goes down or a new kind of error appears, a ticket of the highest priority is raised at once, and the released tickets stay open until it is resolved. A data migration on live data is released only by the person.

The work happens on a development server of our own, set up with Terraform, which holds the projects' development environments. Its sessions can be steered from a phone, through a relay we run ourselves that passes content on only end-to-end encrypted. A question does not stop the process: the agent puts it on the ticket with numbered options and a recommendation, moves the ticket to awaiting information and carries on with other work. The answer comes when the person has time, on the move as well.

  • One rulebook as the only source of the process; only a person can change it
  • No write access for agents to production systems, their data or their credentials
  • Separate roles: building in a sandbox of its own, review by a second AI model, QA by a separate agent
  • QA on dedicated environments with their own database and mail catcher, in a headless browser
  • Merging only by pull request with four green required checks, administrators included
  • Releases within a time window, one at a time, with 30 minutes of watching afterwards
  • A development server of our own, steerable from a phone through an end-to-end encrypted relay

A look inside the product

Recreated view with sample data.

Result

The development server has been running since September 2026, and the process has been in use on an ongoing client project since the beginning of November 2026. There it takes tickets through build, review and QA every working day and ships them to production in several watched releases a day.

The speed comes from the limits, not in spite of them. Because it is settled what an agent may never do, and because the required checks are enforced by the repository itself, the flow stays fast without any limit becoming negotiable. And because every ticket leaves its history on the ticket, from the pick through the review rounds with their verdicts and the QA record with every step to the pull request and the release, every change can be traced back to how it was checked.

What matters to our clients: the rules are not a statement of intent. They sit in a rulebook a person is responsible for, and the merge rule is enforced not by the model but by the repository.

  • Process in use on a client project since November 2026, development server since September 2026
  • Several watched releases per working day
  • Every change traceable on its ticket, with review verdict and QA record
  • Rules, choice of model and data migrations on live data are decided by a person
Collaboration

What clients say about working with us

01 — 03

The client and we as consulting partner were very happy with the quality and the cooperation with Browserbite.

Really feels like working with a partner who cares about the projects as much as we do.

Robert Vossen
Partner, Consulting Agency

Overall, the app has received positive feedback. After Browserbite EOOD implemented the platform, our users found the processes very practical. Previously, we had to provide training for these processes, and the app made everything more intuitive. Our users are happy to have that tool, and we now have better productivity and quality.

They delivered everything on time and on point.

Dennis Goldbach
CEO, DevGold

Within 3 months we executed a live proof of our concepts and reached valuable insights into our business model. We started into the next project phase fast and will further develop our product.

The communication and bilateral understanding were superb.

Dennis Dedaj
CEO, DGTL MKRS

Let’s talk.

30 minutes. Afterwards you’ll know which part of your service can be automated.

Book a call

You’ll talk to Andreas Wawer

+49 176 72944212contact@browserbite.io

Free · 30 minutes

Book a call

In 30 minutes we discuss your project. Afterwards you receive a short written assessment with a suggested scope, a rough budget and timeline, and the recommended next step — whether or not you work with us.

Open the calendar in a new tab instead

Times are shown in your time zone. Prefer to write first? Go to the contact form