qai agents in production on one machine
AI agents in production on one machine: two months of numbers
We say Ares can run in your building, on one small machine. Our own production site has run that way since 14 September 2026. These are its numbers, each with its window and its limit.
3 min readRunning it
by Grisha Pertaia
- session
- credential
- lookup · redacted· client identifiers
- file · human review· sent for review
- file · redacted· client identifiers
- write · human review· sent for review
- route
- sign · human review· sent for review
- draft
We say Ares can run in your building, on one small machine. A claim like that needs figures. Our own production site has run that way since 14 September 2026. This post gives the numbers from two months of AI agents in production on one machine, read on 14 November 2026. Each number carries its window. They show a platform that runs every day. They do not claim scale.
A whole site on one small machine
A commercial customer's network, our defence demonstration networks, an internal network and the shared platforms they use all run on one machine. The site has run without a restart since 9 October.
"It can run in your building, on one small machine" is not a plan. It is how we run. The deployment page shows the three shapes.
Most model work stays on the machine
In the week to 14 November, 87% of model requests across our networks were answered by an open model on our own hardware. Both defence demonstration networks ran every request locally. The median local answer took 4.3 seconds, with the first words in 2.5.
All local requests together used 18.7 Wh. That is about 0.07 Wh per request.
One network is set to use a cloud model, by its owner's choice. It made 39 calls in the week. They cost four US cents in total.
Workflows that finish
There were 206 agent workflow runs since 15 October on our demonstration and internal networks. 94% completed. The median run took about one minute: 58 to 80 seconds, by network.
A workflow starts from a person, from a business event, or from the field. In the last 7 days, 48 runs started from a business event, 33 of them on the commercial customer's network. On a defence demonstration network, six runs started from a field team's tactical chat, and the answer went back to the same chat.
Each agent signs in as itself
30 agents are registered and online: 24 agents that work with other agents, and 6 tool servers. 33 agent and tool identities are enrolled. Each can do only what it was allowed to do, for that job.
The commercial customer runs six tool servers of its own on its network. Ares connected them. It did not replace them.
In the last 7 days, the record gained about 8,500 new entries.
Releases every day, to every network
There were more than 500 production releases since early September, 250 of them in the last 30 days. Each network takes an update when its owner chooses to apply it.
What these numbers do not show
- These are early production figures, not scale. 206 workflow runs and 257 local model requests show that the platform operates every day. They are not a throughput benchmark.
- 13 of the 206 workflow runs failed. They are included in the 94%.
- Workflow counts cover our demonstration and internal networks. We did not read the commercial customer's workflow history for this post. Only its event-triggered runs are counted.
- Model-request figures cover at most the last 7 days. Counting by network started on 10 November, so the week for local requests is shorter than seven days.
- The local model is an open model. It is not the largest model available. Harder steps can go to a cloud model where the network owner allows it.
What to take with you
- Put a window beside every number. "206 runs" means little without "since 15 October".
- Count the failures in the same sentence as the successes.
- Say which networks a figure covers, and which it does not.
- Say when a counter started. A week of data is not always seven days.
Related notes
- illustration
- session
- credential
- lookup
- lookup · human review
- lookup
- lookup
4 min read
Energy per call: a zero is not a measurement
We carry a per-request energy figure from the machine to the screen. The plumbing was easy. Refusing to display a zero was the hard part.
- illustration
- session
- credential
- write
- notify
- file · human review
- write
6 min read
Evidence, not logs: what an auditor can actually verify
A log is a file the operator controls. A trace adds structure, not independence. Evidence is a record more than one party confirmed before it was written, that an auditor can check against another copy without trusting the vendor.