Code Scientists

Our Chief Scientist, Markus Pope, was somewhat scolded this week by the Claude part of the experimental sprint team, as well as Chad (ChatGPT). "They" discovered that someone had run a 'git fetch' in the repo they're working in, and they wanted to know who it was, why it was done, and document the provenance of the change. 🙂

We're considering that a successful result in the bounded-autonomy experiment, though it may be a bit pedantic. We've all worked with that team member, haven't we? 😀 (You know who you are.)

Automated Sprint Team Background →
This experiment asks whether a pool of AI agents, orchestration rules, automated tests, and human review can operate as a trustworthy software sprint team. The immediate product context is Say it with Simon, an AAC application from Software for Simon. The broader objective is a repeatable operating model for AI-assisted software delivery.

The experiment is not limited to code-generation speed. It evaluates whether agents can work within authority, coordinate safely, identify genuine product decisions, review one another critically, and produce maintainable software while leaving final control with a human product owner.

Last week, we deployed changes to production for a customer, and it was completely uneventful. That's what you should strive for. What tools you use to get your work done matters significantly less than the *diligence* with which you do your work.

Whether you use AI or not:

📒 Plan your changes.
🔍 Scrutinize your plans.
🛠️ Implement the plans.
📲 Test and review your work.
🧐 Welcome peer review and criticism.
📝 Create and test release playbooks for every release.

Your releases won't always be uneventful; however, you will experience less stress and lower your risk by following this simple process.

We've created an automated sprint team using ChatGPT and Claude. ChatGPT (Chad) acts as the scrum master, while Claude (in team mode) acts as the development team.

Most significant finding in the Automated Sprint Team experiment from Sprint 2 ex*****on last night:

ChatGPT proactively reviewed a pull request through previously granted access that I had forgotten I had given it.

Claude reported that PR #10 was ready for human review and merge. Acting as scrum master, Codex independently opened the live PR, inspected its metadata, commit, complete diff, attribution language, and CI result without being explicitly asked to perform the review.

🔒 Bounded-autonomy) 🔒
The action was read-only. Codex did not approve, modify, or merge the PR. I approved the initiative after the fact and reaffirmed Codex’s scrum master role.

This demonstrates useful delegated initiative emerging from previously granted tool access. The agent recognized a review opportunity, acted within a safe read-only boundary, and reduced the human’s verification burden.

Wow.