← Blog

Engineering

Meet Tower

· Yoneda Labs · 3 min read

When many agents work in one repository, something has to decide what may land and in what order. We didn't want that something to be another model. Tower is a set of rules on the platform. It reads the state of a Landing and answers, and a recorded state replays to the same answer. It writes no code and starts no agents.

Clearing a Landing

For one exact commit, Tower gives one of three answers, with reasons:

AnswerMeansExample
ClearedIt may landEvery required check and approval passed
HoldWaiting for something that needs no new workAn approval still to come
Go-aroundThe work has to changeA failed check, or a merge that conflicts

The Landing page and the job that lands a Landing ask the same rules, and the job asks again at the moment it lands, so what the page says a Landing is waiting for can't disagree with what the job does. Tower asks no model. An agent's verdict reaches Tower only as a check result or a comment already on record. The Land button and yoneda change work as before and don't go through these rules.

Two holds exist because automation would otherwise land the wrong thing. A Landing from a fork needs an approval of its head, because anyone who can read a repository can fork it, and green checks alone would land a stranger's commit. A repository that asks for no check or approval never lands by itself, or every push would land.

Each decision stays on the Landing with its reasons and the commit it judged. The newest 32 are kept.

Warning about traffic

A session hears about a collision before it happens. When it starts, or reads a file, the platform names any open Landing on the same work:

- Landing 12 ("Retry budget", branch fix) is changing `retry` in src/net.rs

The closest overlap comes first: the same symbol, then the same file, then a Landing that read the file. A warning holds nothing. The session decides what to do with it.

Holding and ordering

If a Landing ahead in line wrote something this one read, or read something it writes, Tower holds the second until the first is out of the way. The normal rules then judge whatever lands second. A Landing that is only waiting holds nobody up, so one left open can't stop a repository.

The line runs queued landings first, then cleared ones with the oldest first, then the rest. Each repository's Tower page shows it: every open Landing, what it's waiting for, its place in line and where it overlaps another.

What we tested

On dev on October 31, two Landings racing on different files both landed. Two racing on the same line gave one landing and one go-around. When 4 jobs raced to land one Landing, it landed once.

Limits

Automatic landing is a per-repository switch, off by default. Where it's off, the Tower page shows what Tower would say and nothing lands on its own. Risk is a beta reading that doesn't hold a Landing, because it isn't calibrated against real outcomes yet. We've tested one Landing at a time and one race of two, so rate at volume is unmeasured. When a legal hold lifts or the switch goes on, Tower looks again at once. When a full repository gets room, Tower hears at its next size reading or compaction, so an idle repository waits for its next event.

More posts

  • · Yoneda Labs

    Introducing Flight Path

    Every change on YonedaRepo gets a route, from the plan an agent files to the record it leaves after landing. What happens at each stop, what we tested and what runs on production today.

    Engineering
  • · Yoneda Labs

    Landings replace the pull request

    When agents write the code, review has to start from evidence. How a Landing works, and what happened when a real agent went through one and nobody pressed Land.

    Engineering
  • · Yoneda Labs

    Does memory help an agent?

    We planted rules an agent could learn only from a memory, then ran it with memory and with a plain clone. With memory it kept the rule 6 times of 6. With a plain clone, 0 of 6.

    Engineering