Axiom's agents draft every release note before anyone is awake

Axiom ships every second Tuesday, and until last year every second Monday was eaten by the same job: turning three weeks of merged pull requests into something a customer would read.

It rotated between two senior engineers. Both of them hated it, both of them were good at it, and the quality visibly fell off in the last third of every set of notes.

What the agent reads

The release agent runs at 06:40 on release Monday. It reads the milestone, the pull requests that actually merged into it, and the last ten published entries — which is where it gets the voice, rather than from a style guide nobody maintains.

It skips reverts, it groups by what a customer can now do rather than by service, and it flags anything it could not describe in one sentence as needing a person. That last behaviour is the one Axiom values most.

It never runs out of care on a Friday. The last three entries are as good as the first three, which was never true when we wrote them.

Ninety seconds each

Two engineers read the draft, usually while the coffee is brewing. The median edit is a sentence, most often to sharpen a claim the agent was too cautious about.

The release email now opens at thirty-four per cent, up from eleven. Replies to it went up roughly six-fold — real questions from real users, which is a category of feedback Axiom did not previously get from release comms at all.

Where it stops

Axiom does not let agents publish. The draft waits for a person, every time, and they have no plans to change that. The saving was never in the clicking.