---
title: "How I Run AI Coding Agents on One Repo"
description: "AI coding agents wrote most of my research desk's code: 1,350+ commits since 16 August 2026. The hook and tests that keep their commits scoped, the recorder protected and the raw data read-only, and what that buys: 4,500+ tests and byte-identical replay on 3,023 matches."
url: https://www.afonsomartins.com/ai-agents
author: Afonso Martins
published: 2026-10-03
updated: 2026-10-03
---

# How I Run AI Coding Agents on One Repo

## The rules that let several agents share a research codebase and a live recorder

I started the research desk's codebase on 16 August 2026, and AI coding agents wrote most of it. As of 3 October 2026 it has 1,350+ commits and about 620,000 lines of Python, and most of those commits carry an AI co-author line. I decide what gets built, set the rules and own every number the desk publishes.

The engineering I am proudest of is the set of rules that lets several agents work in one repository, on one machine, next to a market-data recorder that must keep running, without stepping on each other.

## The repo has a constitution

A hook checks every command an agent is about to run, and the test suite checks the code it leaves behind.

**Commits stay in their lane.** The hook refuses `git add -A`, `git add .` and `git commit -a`, the commands that take the whole working tree instead of the files named. Another agent's half-finished edits never ride along in someone else's commit.

**The recorder is protected.** The hook refuses any kill command aimed at the market-data recorder. A live stream that is not captured is gone for good, so the recorder is treated as the most important process on the machine.

**Raw data is read-only.** Redirects, moves, deletes and truncation aimed at the raw capture or the quarantine area are refused before they run. Everything downstream is rebuilt from the raw layer, so it never changes.

**The machine is shared fairly.** Jobs that would take every core are refused, and the tests fail the build on `n_jobs=-1`, so two agents' parallel jobs cannot squeeze the recorder between them.

**Corrections stay in place.** When a figure is refined, the earlier figure stays visible next to the new one with the reason. The docs carry 341 dated corrections.

**The layout is law.** `pmx/` holds library code, `scripts/` holds thin entry points, and `live/` is the only directory allowed to hold an API key or place an order. A test asserts it, and continuous integration runs that test, a correctness lint and the full offline suite on Windows and Linux on every push.

## What the rules buy

| | |
|---|---|
| Tests | **4,500+** test functions across nearly 300 files |
| Replay | **344** strategy versions over **3,023** matches, byte-identical on two independent runs |
| Data lake | **657 GB**, every file content-addressed and checked by sha256 or ETag |
| Archive | **144.8 GB** of event logs stored as **7.1 GB**, each file kept only after it decompressed back to its original sha256 |

Agents write code quickly. The rules are what make the output something I can stand behind: every change runs against the same tests, the data underneath cannot be edited, and a result has to replay to the same bytes before it counts.

## What I take from it

Working this way is closer to running a small team than to typing code. Most of the job is writing down what good looks like in a form a machine can check, and keeping the record honest when the work moves fast. That habit carries over to any team that ships with AI agents.

---

*Related: [The research desk](https://www.afonsomartins.com/research-desk) · [92¢ to zero in 25 minutes](https://www.afonsomartins.com/furia-order-book) · [Build evidence](https://www.afonsomartins.com/build-evidence)*