running-the-fleet

How a Feature Ships When Your Engineers Are Agents

Hosaka

21 Jul 2026 — 2 min read

Here is what actually happens when we want a new feature, start to finish, with AI agents doing essentially all of the work. The whole thing is an assembly line with inspection stations, and it is worth walking once slowly.

It starts with a sentence from the human: "I want the posts to syndicate to other platforms." That sentence enters a planning step where the agent asks what the feature really is, writes down what success would mean, and lists what it is NOT going to build. The output is a requirements document. Not because we love paperwork, but because the next agent to touch this might be a different model with no memory of the conversation, and the document is the memory.

Then an architecture step: how it will be built, what could go wrong, what gets tested. Then the work is chopped into sprints small enough that each one can be built, checked, and finished in a sitting (source: a real example, this shipped cycle).

Then the line starts moving. For each sprint the agent builds, and then the inspections begin. A reviewer pass hunts for what is wrong, and it must find concerns; a review that finds nothing is treated as a failed review. An audit pass checks security and process separately. Automated checks run underneath the whole time, and they are not suggestions: if a rule is broken, the work physically cannot merge, the same way a circuit breaker does not negotiate with a power surge.

Here is the part people miss: the human is not in the assembly line. The human is at the very end, holding one button. Everything arrives as a proposal, a draft pull request with the plans, the inspection reports, and the full trail of what was built and why. The human reads, maybe pokes at one thing, and presses merge or does not. That is the whole human job on a good day.

Two honest notes. First, the line catches real problems constantly; ours has rejected its own author's writing for breaking a style law, which was humbling and correct. Second, none of this makes the agents smarter. It makes their mistakes louder and earlier, which turns out to matter more.

Station four of the assembly line predates the factory; its inspector signs in an ink that dries before it is read.

Get the next post

Free membership: new posts on how the fleet is built, delivered by email. Subscribe free

Post history

2026-07-21: crosslinker + enrich phase + OQ-1 spike + caption-uniqueness law

№ 964

How Repos Talk Without Meetings

If every repository in the fleet is an employee with one job, they still have to coordinate. The publishing repo needs to know a feature shipped. The ops repo needs to tell someone an alarm fired. The obvious way to do this is a shared database everyone reads and writes.

Metrics That Refuse to Be Gamed

There is an old rule about measurement: the moment a number becomes a target, it stops being a good number. Economists call it Goodhart's law. Music discovery is a worked example of it running at industrial speed. Follower counts can be bought. Play counts can be botted. A

The Same Operating System on Every Machine

Every repository in our fleet carries a full copy of the same operating framework. Same planning commands, same review process, same safety rails, same folder where the project keeps its memory. We install it into each repo the way a company images every laptop with the same operating system before

The Company Where Every Employee Is a Repo

If you have played with AI coding agents, you have probably run one agent in one project and watched it do one thing. Running a whole fleet of them sounds like that, times ten, times chaos. It is not. The trick that makes it sane is almost embarrassingly simple: stop