The original 10x engineer

Everyone is using AI to code, everyone is going faster, but how much faster?

We are seeing a 4x increase of productivity in the last 6 months:

This did not come for free though, it required adapting to a new mindset and deeply changing how we produce code. Let’s dive into how software is built at Comand AI.

The first building block

The longer AI can run unsupervised, the most efficient it is. Switching between agents & projects has a productivity and mental cost. One of our early goal was to allow the agent to produce a full PR unsupervised, which would require minimal time for a human to review and little to no iteration.

To get there, the agent runs with a set of shared guidelines, Comand AI harness. It will:

Being able to consistently build a high-quality PR that will require little effort to be merged is the foundation for the next stage.

Accelerate

Best teams iterate faster than the competition. You can’t expect extraordinary results using ordinary methods: it’s time to think outside the box.

The next bottleneck becomes the code-review. When our retro came up and I wanted to propose to skip them for the majority of PRs, I was a bit worried of how it would be received by my peers. To my surprised, bypassing the review step was mentioned by the majority of engineers before I said anything!

When taking a step back, it makes total sense: given the amount of AI-produced code, it does not scale to review everything with human eyes. Engineers understood that and were using agents to assist reviews anyway. So we decided to integrate this in the PR process: we have already the self-correct review, so we just need a bot to ensure that:

• the PR building process has run all steps

• the risk is acceptable

If both are true, our bot will approve the PR. This means only PRs touching critical features, architecture and infra get reviewed, giving a lot of ownership to engineers.

Once this was in place, the slowdown became our CI: a 10-15 minutes CI becomes painful when code changes so quickly. We optimized ours to run in 3 minutes. Conveniently, AI is great at this.

Embrace entropy

Going this path where everything evolves much faster through AI generated code creates entropy.

We needed ways to keep the entropy over control. So we introduced agent bots.

Bots that run nightly to ensure consistency and open refactor PRs.

Bots that tracks what could be improved in our harness to reduce noise & iterations.

Bots that handle flaky tests and open fix PRs.

Bots that monitor our CI and investigate when it’s too slow.

Bots that listen to production logs, investigate and open PRs when needed.

The Software Factory

When all the above systems are put together, it becomes a Software Factory.

Creating a change just requires to speak a few sentences to an agent, and you come back 15-30 minutes later to a PR ready to merge. It’s in production 1h after you had the idea. You spent less than 5 minutes on it.

The automated production with feedback from several quality assurance systems is widely used in traditional factories. When it starts working for software the efficiency gains are compounding.

AI is now better and faster then most engineers at writing code, myself included. But the software engineer role is not only about that. Our job moved into architecture, product, prioritization, and most importantly: running the factory.

Join us in the Software Factory and make a difference in the defense of Europe: https://www.comand.ai/careers

By Raphaël Boucher, Software Engineer at Comand AI