A crew on your problem
- Who does the work
- A team of engineers and designers, with a lead.
- Rhythm
- Sprints and demos.
- What you pay for
- The team's time.
- What makes it cost more
- More people, for longer.
A Machinus engagement is not a regular CONFLICT engagement. A regular one puts a crew on your problem. A Machinus one puts the factory on it, with a machinist running the line. That changes who does the work, how fast it moves, and what you pay for.
A machinist is a person at CONFLICT who runs your line: they approve plans, sign off what ships, unblock the agents and talk to you. The factory builds. Your machinist runs the line. You choose the coverage.
One machinist covers your working hours and shares their day across a few engagements. The factory keeps working overnight and queues what needs a person.
Machinists in different time zones hand the line to each other, so there is always a person awake to approve, sign off and answer. Three people keep a line fully unblocked around the clock.
A team on every shift, for high volume or work that cannot wait. More lines running at once, more hands to sign off.
This is not cheap. Thousands of anything never is. A factory is about speed, consistency and quality, at a size where those are hard to get any other way. This section explains how the bill is built. The numbers are other companies' published figures, linked to where they came from. They are not ours.
At Uber, more than 70% of pull requests are attributed to agents. Across 22,000 developers in another study, median time in review went up 441%. Writing stopped being the slow part. Uber Faros AI
One measured set of 18 agent-built changes averaged about 234 million tokens of context and 700 round trips to the model each, for about $65 a change. Almost none of that is the code it wrote. Sonar
Across 12,000 developers, token cost per merged change ran from $0.28 to $89.32. That is why usage is metered per run and shown to you, not guessed up front. Jellyfish
The same study found nearly ten times the tokens bought about twice the output. Throwing compute at a problem is the expensive way to be slow.
In a study of 8,106 agent-written fixes, 26.1% were closed without ever merging. A factory's job is to make that number small, and to stop early when a run is going nowhere. Alam, Mondal and Roy
Uber grew agent requests 9.4 times in six months while spend levelled off and cost per session fell 52%, by routing work to the right model and cutting wasted context. The same discipline is what makes a line fast and consistent, and it is what the flat rate pays for.
So they are the part you choose. The line runs without a crew. A machinist approves, signs off and answers for it, at the coverage you pick.