Blog · How we work

How we build software with AI agents: who decides, what the machine does and what a person reviews

By Pedro Mulero Sánchez, CEO and founder of Tarraco App Lab · 07/07/2026 · 7 min read

Almost all the code we write today is typed by an AI agent. That doesn't mean the agent decides. Here is how we split the work between the machine and a person, with the numbers from our own repositories.

In a nutshell

Every project has a rules file written by Pedro Mulero. The agent reads it before touching anything and it spells out what the agent may not do, starting with publishing.

The agent writes the code, the tests and the message for each change. In Perito Tacho it is listed as co-author of 394 of the 395 commits as of 11/11/2026.

Nothing goes live until a person has reviewed it locally. A prototype takes us days. What takes weeks is deciding the rules and checking them.

This article is accurate as of: 11/11/2026.

Who decides: a person and written rules

The AI agent doesn't choose what gets built or how the product behaves. Pedro Mulero decides and writes it down in a file that lives inside each repository, the AGENTS.md. It's an open format that several coding agents understand: a README written for the machine.

The rules aren't about style. They are the ones that cause real damage if broken. A few, as they appear in our repositories:

  • On the Tarraco App Lab website: everything stated about the products has to be true in the products.
  • In Perito Tacho, the tool used to prepare expert witness reports: not a single outgoing request while it runs, because personal data and evidence go through it.
  • Also in Perito Tacho: the same input file must produce the same output, byte for byte. A report that can't be reproduced can't be defended in court.
  • One more: every figure in a report must be traceable to the file and the record it comes from. If it has no source, it doesn't get printed.

Every new rule records who set it and when, for example «Fijado por Pedro el 04/10/2026» (set by Pedro on 04/10/2026). Months later everyone knows why it's there and nobody removes it for convenience.

What the AI does

The agent writes the code, writes the tests, runs them and drafts the message for each change. We work with Claude, by Anthropic. Every commit it takes part in says so in its footer («Co-Authored-By: Claude»), so it can be counted:

ProjectFirst commitCommits as of 11/11/2026With the agent as co-author
Tarraco App Lab website and ERP18/06/2026247239
Perito Tacho24/07/2026395394
Driver Solution18/05/2026234195

The agent writing the code doesn't mean it works alone. Every session starts from a specific request and ends with the change made on the local machine, unpublished. If something can't be checked, we say so. We don't fill the gap with whatever seems likely.

What a person reviews

There is one line no agent crosses: publishing. On the Tarraco App Lab website, no agent runs the deployment commands and Pedro deploys after reviewing locally. In Perito Tacho, no agent publishes or installs anything.

Before anything goes out, a person reviews three things:

  • The result, locally. The website is checked on the development server, just like in production. Opening the HTML with a double click gives a broken page that proves nothing.
  • What is claimed. Figures, dates and legislation. A legal citation is checked against the Spanish Official Gazette (BOE) or the Official Journal of the EU before it is written, never from memory.
  • What gets signed. Perito Tacho documents are signed by the expert and go to court. The agent prepares the draft and the person reads all of it before issuing it.

Tests: every rule has its own

A rule nobody checks ends up forgotten. That's why many rules in Perito Tacho end the same way: «Lo prueba test/…» (tested by test/…). The clean-up of the case register is guarded by test/depura-libro.test.js. Document layout, by test/maquetacion.test.js.

ProjectTest files
Perito Tacho143
Driver Solution31
Tarraco App Lab website and ERP11

On the Tarraco App Lab website the tests focus on the ERP: the invoice hash, roles, access control and bank reconciliation. Perito Tacho has no server and no continuous integration: tests run on the machine itself with node --test. Its calculation core doesn't touch the disk or the network and doesn't read the system clock. So the same file always gives the same result and a test can prove it.

Why a prototype takes us days

With an agent, writing code is no longer the slow part. Perito Tacho's history shows it: the first commit is from 24/07/2026 and by 27/07/2026 it already had a chained case register and launched as a Windows application. There were 14 commits that day.

What takes time is what comes next: deciding the rules, testing them with real data and fixing what a person sees and the machine doesn't. Since that first commit Perito Tacho has reached 395. A working prototype is the start of the job, not the end.

What we have learned

If you work with AI agents, this is what has worked for us:

  • Write the rules down before you need to remember them. Whatever isn't in the rules file doesn't exist for the agent in the next session.
  • Put a date and a name on every rule. A rule with no author ends up questioned or deleted.
  • Turn the important rules into tests. A test raises the alarm on its own. A paragraph doesn't.
  • Keep for a person whatever can't be undone: publishing, signing and sending.

You can see what we have built this way in our case studies. If you have any questions about how we work, write to us from the contact page.

Frequently asked questions

Is the code written by an AI?

Mostly, yes. In our three main repositories the agent is listed as co-author of the vast majority of commits. Decisions and review belong to a person.

Can an agent publish something on its own?

No. The rules files in our projects expressly forbid it: Pedro deploys after reviewing locally.

How do you know what the agent does works?

With automated tests run on the machine (143 files in Perito Tacho, 31 in Driver Solution and 11 in the Tarraco App Lab website and ERP) and a human review of the result.

What about the personal data each program handles?

It's decided before a single line is written and recorded in the project rules. In Perito Tacho, for example, no data leaves the machine except for three exceptions authorised in writing, each with its own scope.

What is an AGENTS.md?

A text file inside the project with instructions for the agent: which commands to use, where everything is and what it must never do. It's an open format.

Any questions about this?

Write to us from the contact form and a member of the team will get back to you.

Sources