How we use AI building Rocket: agents build, people decide
Rocket is built with coding agents every day. Here is what we hand to them, what we never do, and why.
Part of our series on taste and shipping fast.
We practise what this series preaches. Most of Rocket's code is written with coding agents, and changes reach our live demo many times a day. That speed is real. It works because the decisions stay with people.
What the agents do
Write the code for a slice once it is decided, following the repository's own rules and skills.
Write the tests that prove it, and run the full local CI before anything merges.
Review each other's work with fresh eyes, and run audits (security, compliance, code quality) that check the code against written standards.
What stays with people
What to build, for whom, and what to leave out.
Every decision is written down with its reason, so the next person (or agent) can see why.
Changing an existing test needs a person's approval, because a test is a promise and agents shouldn't quietly change promises.
Releases to products built on Rocket happen only when a person says yes.
Why it works
Agents are fast and tireless, and they will build exactly what you ask for, including the wrong thing. The guard rails above are how we keep the speed and lose the risk. The taste part, deciding what is worth building, is the one job we haven't handed over, and we don't plan to.
If you start your product on Rocket, you get the same setup: the rules, skills and checks for coding agents come with the platform, so your agents build fast and your people stay in charge of what matters.