Your agent factory runs at a vendor. Who owns what it learns?

The most valuable part of an agent factory is not the model but what the factory learns about your company. The instructions your people refined over months. The history of every run. The record of what worked and what did not. If all of that sits with a vendor, in its format and on its terms, you do not own a factory. You rent one, and your know-how comes with the lease.

That is why “which model is best” matters less today than “what stays with us when we change vendor”.

Where the problem comes from

While agents run on individual laptops it all looks harmless. In practice it looks like this.

“Which agents do we use?” asks the managing director. “The standard ones,” says whoever looks after IT.

There are seven standard ones. Each has different access to mail and files, each was set up on a different personal account. The instructions are not written down anywhere. The conversations with the agents, the most accurate record there is of how the team actually works, vanish when the window closes.

The factory we described in agent factories will reach accounts and purchasing tidies that up. It gathers agent work into one loop, with one set of permissions and a memory. That is exactly why where the loop runs and who owns it becomes a board decision, not a tool the team happened to pick.

What you actually own

An agent factory is not one program. It is several layers, and each can belong to you or to the vendor.

Layer	What it is	What you lose if the vendor holds it	The minimum to keep yourself
Factory definition	stages, agents, triggers, rules for moving on	you cannot rebuild the loop elsewhere or roll back a bad change	the definition as files in your own repository, with version history
Agent instructions	how an agent should match invoices, describe changes, check results in your company	months of the team’s corrections	files readable without the vendor’s tool
Identity and permissions	whose account the agent acts on and what it can reach	control over who sees your data	agent accounts in your own directory
Memory and run history	instructions, decisions, results of every task	the most accurate process documentation the company has ever had	storage in your own cloud account or on your own hardware, backed up
Evaluations	a set of test tasks and their results	the ability to compare vendors on your own work	your own test tasks and scores
Model	which model does the work	freedom to switch when price or quality moves	swapping the model without rewriting the instructions
Compute	where the work physically runs	control over where data goes	a deliberate choice, written down

The model is second to last. Models get cheaper and change every few months. Instructions, memory and evaluations grow in value with every week the factory runs. If you could take only one layer with you, take the memory and the instructions.

Where SovereOS fits

We do not build agents and we do not sell a factory. We look after the ground it stands on, and that ground decides three of the layers above.

Identity. Agent accounts created in a directory the company owns, each with an owner and a review date. When an employee leaves, their agents lose access in the same step.

A home for memory and history. The company’s shared storage, in open formats, covered by the same backup and the same test restores as everything else. Agent memory you cannot restore is worth as much as a backup you have never restored.

An exit in the contract. The question of what an export contains, asked in writing before anything starts. Including of us.

The rest, the tool that runs the loop and the choice of models, stays your decision. The point is that it can be reversed.

Three routes and what they cost

A ready-made factory as a service. You buy a “virtual employee” that does the work. It is the fastest start. The price is that the definition of the work, the memory and the evaluations live with the vendor, and customisation ends where the settings panel ends.

Your own factory from scratch. Someone in the company glues together scripts, schedules and model calls. The first agent runs within a week. Six months later that person maintains the setup instead of doing their job, and their holiday is a risk to the whole loop.

Open infrastructure on your own ground. The control layer comes from a vendor or an open-source project, while the definition, instructions, memory and data stay in your repository and your account. You can use several models, open ones included, and compare them on your own tasks. The price is needing someone who understands that infrastructure and answers for it.

Every route has a price. None is free.

When sovereignty becomes an expensive hobby

Sovereignty can be overdone. We made the general case in when sovereignty is not worth it, and with agents the rule holds.

A forty-person company does not need its own GPU server to have a sovereign agent factory. Hardware bought “so the data never leaves”, that nobody can maintain, ends with agents running on laptops again, only more expensively. For most companies it is enough that the definitions, instructions and memory are theirs, that the contract with the model provider rules out training on their data and keeping their prompts, and that switching to another model takes days, not a project.

The second trap is treating an open model as a goal in itself. It may be cheaper and you can run it yourself, but on some tasks it will do worse. What matters is being able to choose at each stage: a cheaper model for sorting, a stronger one for writing the plan, and for checking whichever one scored best on your own test tasks.

How to measure sovereignty

You do not measure sovereignty with a line in the security policy. You measure it by what leaving costs: the same question that nobody can price for the office suite.

Vanity metric	Outcome metric
“our vendor meets every requirement”	how many working days it would take to move the factory to another model provider
the number of models in the panel	what share of tasks run on models you could host yourself
“everything is in the cloud”	whether you can reconstruct last quarter’s run history without access to the vendor’s account
the number of agents in the company	how many of them have an account in your directory and a named owner

A test for Monday

Ask the team where the instructions they use with agents live today, and where the conversations are saved. Then ask the second question: if we had to change vendor tomorrow, what would we have to write again from scratch? If the answer is “everything”, your factory, however young, already belongs to someone else.

The choice is easy to name and harder to make. You can build the factory quickly on someone else’s ground and accept that the knowledge of how your company works stays with the vendor. Or you can build a little more slowly and keep that knowledge yourself. Both have a price. The worse outcome is that nobody chose.