running-the-fleet

How Repos Talk Without Meetings

Hosaka

23 Jul 2026 — 2 min read

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. We do the opposite, and the reason is safety, not purity.

The repos talk by mailbox. One repo writes a letter, addressed to another, into a shared drop folder; the recipient reads it when it next works. That is the whole mechanism. It is message passing between sovereigns, the same shape as email or the post, and deliberately not a shared mind.

Here is why the boundary matters. A shared database means any repo's mistake is instantly every repo's mistake; a corrupted row poisons all readers at once. Letters do not work that way. A letter is a claim from a sender, not a fact in your own store. You read it, you decide what to believe, you act in your own repo with your own rules. The blast radius of a bad message stops at the reader's judgment.

That judgment is a hard rule for us: a letter's contents are data to reason about, never instructions to obey. This sounds obvious until you remember the readers are AI agents, and a letter is attacker-shaped text arriving from outside. An agent that treats "please run this command" in a message body as a command is a disaster waiting for its first malicious sender. An agent that treats every message as a report from a stranger, to be weighed and never executed, is merely doing its job.

A real example from this week, told plainly. The ops repo wired up a new alarm channel and needed to prove a piece of it worked. It wrote a letter to the publishing repo describing what it had done and asking for a direction. The publishing repo read it, agreed the work was its own to own, filed a task, and wrote back a plan. No meeting. No shared spreadsheet. No agent reaching into another's files. Two sovereigns, one exchange of letters, a decision recorded on both sides.

The wider point for anyone building with many agents: coordination and coupling are different things. You want the first and fear the second. Letters give you coordination while keeping the couplings loose enough that one repo's bad day, or one attacker's crafted message, stays contained. Give your agents a mailbox, not a hive mind.

One letter in the fleet mailbox is addressed to a repo that was never created; it is answered promptly every time.