Spoiler: Our Android developers don't open Android Studio everyday!

The Android team at Photoroom was one of the first teams here to experiment with AI and agentic workflows. It started with individuals, then spread to the whole team, then across teams, and now anyone in the company can open a pull request against our Android repo.

This talk is for anyone who wants to set sail and solve "coding" in their own organisation. You get the full set of experiments we ran, how the journey started, where we landed, and what we learned on the way, including the parts that did not work.

The short version: like every other developer, we started with Claude or Codex inside our own IDEs. Then came parallel Worktree setups, then our own ADEs. Here is what we had to solve along the way, and what you can take from each of it.

1. Adapting the harnesses for high quality results, and making TDD the obvious choice
Tests turned out to be the cheapest way to keep an agent honest, so TDD stopped being a discipline argument and became the default.

2. Building the right skills, and a process for the team to create and improve them
A skill is only worth writing if someone else on the team will reach for it, so we built a process for proposing, reviewing and retiring them.

3. Building the knowledge base and compound engineering
Every solved problem should make the next one cheaper, which only happens if the agents can read what you already learned.

4. Left first approach, a mini knowledge base
Writing the tickets and speccing each one before any execution is what removes most of the intervention later.

5. Reviewing PRs, and making Claude and Codex verify each other's work
Two models with different blind spots catch more than one model reviewing itself.

6. Remote agent instances, like Coder
Remote instances are excellent for some jobs and frustrating for others, and knowing which is which saves a lot of wasted setup.

7. Measuring agentic AI assisted PRs
If you cannot say what improved, you cannot defend the investment or spot where it is quietly getting worse.

8. Slack agents, from creating a ticket to starting an investigation or opening a PR
Moving the entry point into Slack is what let people outside the Android team contribute to the Android repo.

9. What's next for us
The open problems we have not solved yet, and where we think this goes..