---
title: "Who owns what your agent factory learns?"
url: "https://predictes.com/blog/who-owns-what-your-agent-factory-learns"
description: "An agent factory collects knowledge about how your company works. Before choosing a supplier, check what stays with you when you want to leave."
---

# Who owns what your agent factory learns?

November 4, 2026·5 min read·predictes

The most valuable part of an agent factory is not the model but what the factory learns about your company. The instructions your people corrected for months, the record of every run, the quality scores, the memory of what worked and what did not. If all of that sits with a supplier, in the supplier’s format and on the supplier’s terms, you do not own a factory. You rent one, and your own know-how comes with the lease.

That is why “which model is best” now matters less than “what stays with us when we change supplier”. In [the first part of this series](https://predictes.com/blog/agent-factories-will-not-stay-in-engineering) we argued that the factory will spread from engineering to the rest of the company. The wider it spreads, the more this question weighs.

## Where the problem comes from

While agents run on individual laptops, ownership looks harmless. In practice the picture is familiar: every team installs a different tool and gives it broad access. Instructions and connected tools are written down nowhere. Conversations with agents — the most detailed record of how your teams actually work — disappear when the window closes. Nobody can tell whether work with agents is getting better, because there is nothing to measure it from.

A factory fixes that. It pulls agent work into one loop, with one set of permissions, with metrics and memory. Which is exactly why the question of where that loop runs, and who owns it, becomes a board decision rather than a tooling choice made by one team.

## What you actually own when you have a factory

An agent factory is not one program. It is a stack of layers, and each can belong to you or to the supplier.

Layer

What you lose if the supplier holds it

The minimum worth keeping

Factory definition: stages, agents, triggers, rules

You cannot rebuild the loop elsewhere or roll back a bad change

Definition as code in your repository, with history

Agent instructions and skills

Months of corrections your teams made

Plain files in your repository, readable without the supplier’s tool

Tools and permissions

Control over who sees your data

One permission register that you manage

Memory and run history

The most accurate process documentation you will ever have

Stored in your cloud tenant or your data centre

Evaluations and metrics

The ability to compare suppliers on your own tasks

Your own test set and results

Model and runtime

Freedom to switch when price or quality changes

Swapping the model without rewriting instructions

Compute

Control over where the data is processed

A choice: your own, cloud with no retention, or mixed

Notice that the model sits near the bottom. Models get cheaper and change every few months, and a model being retired is a routine event, as we described in [the vendor retired your model](https://predictes.com/blog/the-vendor-retired-your-model-what-happens-to-your-system). Instructions, memory and evaluations gain value with every week the factory runs. If you could take only one layer with you, take the memory and the instructions.

## Three routes and what each costs

Rented

-   –Factory as a service: fastest start
-   –Definition, memory and scores stay with the supplier
-   –Customisation ends where the settings panel ends

Owned

-   →Open infrastructure in your environment
-   →Definition, instructions, memory and data in your repository and tenant
-   →Needs a named person who understands and runs it

The third route, building everything from scratch, is owned too, but the price is paid in plumbing.

**A factory as a service.** You buy a “virtual team member” that does the work. The start is the fastest. The price is that the definition of the work, the memory and the scores stay with the supplier, and adapting it ends where the settings panel ends.

**Your own factory from scratch.** The team glues together scripts, queues and model calls. The first agent works within a week. The price arrives later: after half a year your best engineers maintain plumbing instead of the product, and every new stage of the loop is a new project.

**Open infrastructure in your environment.** A supplier provides the control layer: the work queue, the dashboard, metrics, a way to run agents. The factory definition, the instructions, the memory and the data stay in your repository and your cloud. You can use several models, open ones included, and compare them on your own tasks. The price is that somebody in your organisation must understand that infrastructure and answer for it.

Before signing for any of the three, ask what you would receive on exit, in which format, and how long it would take. The same question works for any AI application, see [before anyone builds your AI application, agree whose it will be](https://predictes.com/blog/before-anyone-builds-your-ai-application-agree-whose-it-will-be), and for the run history in particular, [ask an agent supplier for the decision log, not the demo](https://predictes.com/blog/ask-an-agent-supplier-for-the-decision-log-not-the-demo).

## Where sovereignty becomes an expensive hobby

Sovereignty can be overdone, and it is worth saying so.

A company of eighty people does not need its own GPU servers to own its agent factory. Buying hardware “so the data never leaves”, when nobody can maintain it, usually ends with agents running on laptops as before, only at a higher cost. For most companies it is enough that the definitions, instructions and memory are theirs, that the contract with the model provider excludes training on their data and retaining prompts, and that the architecture can move to another model within a week.

The second trap is treating open models as a goal in themselves. An open model is often cheaper and can run in your own environment, but on some tasks it will be weaker. The factory makes sense when each stage can use the model that fits it: a cheap one for sorting, a strong one for writing descriptions, and for review the one that scores best on your own test set. That test set is the part to build first — [build the evaluation harness first](https://predictes.com/blog/build-the-evaluation-harness-first).

## How to measure it

Sovereignty is not measured by a sentence in a security policy. It is measured by what leaving costs.

Vanity metric

Outcome metric

“Our supplier is compliant with every regulation”

Working days needed to move the factory to another model provider

Number of models available in the panel

Share of tasks running on models you could run yourselves

“Everything is in the cloud”

Whether you can reconstruct last quarter’s run history without the supplier’s account

A useful parallel is the data layer. An open table format is worth choosing because the exit plan can rely on it, as we wrote in [Apache Iceberg: the table format your exit plan can rely on](https://predictes.com/blog/apache-iceberg-the-table-format-your-exit-plan-can-rely-on). The factory’s memory deserves the same treatment as the company’s tables.

## Our part

In [AI strategy](https://predictes.com/services/ai-strategy) we help a board decide which layers must stay in-house and which can safely run with a supplier. We build on the cloud you already use or on an open-source stack you own outright, and we will tell you when the platform you already pay for is enough. More about predictes is on the [home page](https://predictes.com/).

## A test for Monday

Ask your teams two questions. Where do the instructions they use with agents live today, and where are their conversations stored? Then the second: if we had to change model provider tomorrow, what would have to be rewritten? If the answer is “everything”, your factory, however young, already belongs to someone else.

The choice is easy to name and hard to make. You can build the factory quickly on someone else’s infrastructure and accept that the knowledge of how your company works will sit with them. Or you can build it a little more slowly and keep that knowledge at home. Both have a price. The worse outcome is making neither choice deliberately.

-   agent factory
-   vendor lock-in
-   open source
-   exit plan
-   AI agents

## Keep reading

-   [One source, three stages: detecting an autonomous attack](https://predictes.com/blog/one-source-three-stages-detecting-an-autonomous-attack)
-   [How to build an agent factory: the loop, stage by stage](https://predictes.com/blog/how-to-build-an-agent-factory-the-loop-stage-by-stage)