Engineering agents with their own machines

Stop shaving yaks.Hire one.

Other agents hand you a pull request to check. A yak carries the ticket further: it runs your real stack, fixes its review until it's approved, and hands it to you to merge. Then it watches production.

Sign in with Google. $20 of credit to start, $180 more when you connect GitHub. No card.

A large black yak standing calmly in the aisle of a bright software office while engineers keep working at their desks.

One real ticket, start to finish

It doesn't stop atthe pull request.

YAG-1368 on our own repository, from the moment a yak picked it up to the moment it was live. Times are hours and minutes after the hand-off.

  1. 0:00

  2. 1:06

  3. 1:08

  4. 1:57

  5. 2:06

  6. 3:08

  7. +10 min

    After every deploy

    Then it watches production

    Ten minutes after the deploy, the yak reads your logs and error tracker for anything new and says in the ticket what it found. If an error looks like its own, the fix is one click away: a new ticket the yak takes, and a person merges that too. It never deploys and never rolls back on its own.

    This is live in production.

    No new errors in the 10 minutes since this went live. Checked Datadog logs and Sentry.

    What the yak writes in the ticket

Times come from the ticket, the pull request and the deploy log. Our repository is private, so these are excerpts.

We build YAGNI with yaks. These are real pull requests they wrote on this product, merged by one of us.

  • A yak's "Needs you" question drops internal ids and plan jargon#3235 · merged Nov 18
  • The thread shows a live line while the yak plans#3222 · merged Nov 18
  • A split child's PR names the files outside its piece#3223 · merged Nov 17
  • Start a yak's setup once running copies are on#3215 · merged Nov 17
  • Yak errors report under their own Sentry area#3214 · merged Nov 17
  • A check on a yak means clicking it in the browser#3211 · merged Nov 17
  • PR ready for review makes no new chat message#3197 · merged Nov 17
  • A yak keeps fixing review comments until the review approves#3196 · merged Nov 16
  • A weekly count of clean yak deliveries, with stops by cause#3072 · merged Nov 10
  • A failing test or an empty search never marks a turn as failed#3069 · merged Nov 10
  • The Work list shows the tracker key a pull request is titled by#3046 · merged Nov 9

Let the yaks dothe yak shaving.

Every engineer knows yak shaving: the setup, the flaky test, the upgrade, the review comment that stands between you and the work that matters. We give your team yaks for all of it.

A yak standing in morning mist, looking at the camera.

yak /yak/ noun

  1. A sure-footed animal that carries heavy loads, steadily, through bad weather.
  2. An engineering agent with its own machine and your stack running on it. It takes tickets from your tracker and brings back pull requests that were built, reviewed and tested, for a person on your team to merge. Then it watches the change in production.
  • Its own machineNot your laptop, and not a throwaway sandbox.
  • Your stack, runningSet up once with you, and kept between tickets.
  • Asks only when it mustYou hear from it when a decision is yours to make.

How it works

From sign-in to production,on your real system.

  1. 02

    It gets your stack running

    The yak sets your system up on its own machine, with you, and keeps it running. It asks for the few things it cannot find, such as a token, and shows you every command it ran.

    its machine, not your laptop

Every ticket is

Built
against your running system, not a guess at it
Reviewed
and fixed, round after round, until the review approves
Tested
in a real browser, when you turn QA on
Merged
by a person on your team. That part stays yours.
Watched
in production after the deploy, for anything it broke

Not a sandbox.Your whole stack.

We set up a working copy of your development environment with your team, and check it works before a single ticket runs. Your yaks see what your engineers see.

A throwaway sandbox

One repositoryUnit tests

Gone when the task ends.

Where your yaks work

The app, runningYour databaseQueues and workersEvery repositoryYour internal CLIsThird-party APIsA real browserLogs and traces

Kept, checked and updated. Every ticket starts from it.

Volume

One of you.A herd of them.

Every ticket gets its own machine, so several run side by side instead of one after another. Hand out a morning's worth of tickets, go do the work only you can do, and come back to pull requests.

  • Several at once. Three tickets run side by side to start, and we raise that when you need more.
  • No babysitting. A yak asks when it needs a decision and keeps going when it does not.
  • No seats. Add a yak for a busy week. Let it rest after.

4 tickets in flight1 of you

Four yaks. Standup got shorter.

Pricing

Pay for the minutes a yak works.Not the hours it waits on you.

One ticket, start to merge

You are never billed for a yak that is waiting on you.

Fair questions.

What is a yak?

An engineering agent with a machine of its own and your stack running on it. You hand it tickets from your tracker. It plans, builds, runs your tests, answers its code review, and opens a pull request for a person to merge. After the deploy, it watches production.

Why a yak?

Yak shaving is the pile of small jobs standing between you and the thing you sat down to do. A yak carries heavy loads, steadily, in bad weather. It seemed like the right animal for the job.

How do I start?

Sign in with Google and your workspace is ready, with $20 of credit. Connect GitHub (that adds $180), choose the repositories, and hire your yak. Its first job is getting your system running on its own machine, and it asks you for anything it cannot find. If you would rather do it with us, book 30 minutes.

How is this different from the assistant in my editor?

An editor assistant needs you in the chair. A yak has its own machine, so several can work several tickets while you are in a meeting. You come back to pull requests.

Does a yak merge its own work?

No. A yak opens a pull request, the review reads it, QA can test it in a browser, and a person on your team merges it.

What happens after I merge?

The yak follows its change out. Ten minutes after the deploy, it reads your logs and error tracker (Datadog and Sentry today) and says in the ticket what it found. If an error looks like its own, the fix is one click away: a new ticket the yak takes, for a person to merge. It never deploys or rolls back on its own.

Will it work on our stack?

A yak's first job is getting your system running on its own machine, whatever it is written in. It asks for the few things it cannot find, such as a token, and it tells you plainly when something will not boot.

What does it cost?

You pay for the minutes a yak spends working, plus the model's tokens at the provider's list price and a 7% processing fee. Waiting on you, on a review or on CI is free. Checking production after a deploy costs only its tokens. There are no seats.

What can a yak reach?

The repositories you pick, through the YAGNI GitHub App, the tracker you connect, and the logs and error tracker you connect for watching production. It never signs in to your production systems. Each yak works on an isolated machine. The security page has the detail.Read the security page.

An engineer leans back in a desk chair with hands behind their head while a black yak lies asleep on the carpet by the window.

Hand it over.Breathe out.

Give your yak the ticket you have been putting off since spring. It will tell you when it's live.

Sign in with Google. $20 of credit to start, $180 more when you connect GitHub. No card.