# Self-hosted AI agents: who runs the control plane

> Self-hosted AI agents put the control plane, session data, models, and updates under your team. Compare those responsibilities with a managed deployment first.

URL: https://factory.com/articles/self-hosted-ai-agents
Published: 2026-09-18
Updated: 2026-10-23

Self-hosted AI agents run their control plane, the service that coordinates sessions and supporting services, inside the customer's own infrastructure. That puts the service under the customer's control, and it also makes its updates, availability, and recovery part of the customer's operating work. Installing a coding agent locally is a different job from operating the service behind its sessions.

Before choosing customer hosting, identify the team that would restore the control plane after a failed upgrade and the process it would use. An internal model key or a local executable cannot answer that operational question.

## Self-hosted AI agents change who operates the system

Factory offers a managed control plane and a customer-owned alternative. [Factory Managed](https://factory.com/deployment/factory-managed) provides hosted orchestration and operations while Droids can run across existing development environments.

[Factory Private](https://factory.com/deployment/factory-private) puts the control plane in customer infrastructure. The published comparison identifies differences in ownership, data boundary, and access controls.

The ownership choice sets the operator's responsibilities. Model routing and execution permissions still need their own configuration, whichever control plane the organization uses.

## Review self-hosted AI agents beyond the model key

Bringing an approved model does not move every supporting service into the customer network. Factory's sovereign software development whitepaper distinguishes Managed control-plane and session-data hosting from Private's customer-hosted placement.

Review session handling separately from inference. Record where context is stored, which model service receives requests, and what the gateway logs. Include retries, fallbacks, and support diagnostics in the diagram.

Factory's [deployment documentation](https://docs.factory.com/enterprise/network-and-deployment) describes the runtime and traffic differences between cloud-managed, hybrid, and airgapped patterns. Use the intended pattern to test the complete workflow, not just its first successful model response.

## Budget for the responsibilities that move

Operating a private control plane requires a release and recovery process. For example, rehearse an update in a disposable environment, then restore the previous approved configuration and run a known task. A successful recovery should not depend on undocumented steps in one administrator's shell.

The model service needs an owner too. A new model version can alter task behavior even when the agent build stays fixed. Keep representative validation tasks and review the results before promotion.

Monitor the dependency that can stop work. Adding execution machines will not repair a saturated inference service or an unavailable internal package mirror. Distinguish those failures from agent errors so the response goes to the team that can resolve them.

Support access also belongs in the plan. Agree on which diagnostics may leave the environment, how sensitive content is removed, and who approves release. Avoid making troubleshooting an undocumented exception to the boundary.

## Treat disconnected operation as its own requirement

Airgapped operation adds a no-runtime-connectivity requirement. Factory's [airgap build](https://docs.factory.com/enterprise/airgapped-deployment) uses customer-configured models and disables Factory-bound services. Cloud-dependent features such as Slack integration and hosted analytics are unavailable.

Before developers depend on the service, make sure another operator can promote an update, recover a failed one, and collect permitted diagnostics. Those procedures determine whether customer control is usable when the system needs maintenance.

## Frequently asked questions

### What are self-hosted AI agents?

Self-hosted AI agents run their control plane, the service that coordinates sessions and supporting services, in customer infrastructure. With Factory, Factory Private provides that customer-owned option, while Factory Managed provides hosted orchestration and operations.

### What changes when an AI agent control plane is self-hosted?

Factory Private puts the control plane in customer infrastructure, which makes its updates, availability, and recovery part of the customer operating work.

### Does bringing an approved model move every service into the network?

No. The sovereign software development whitepaper distinguishes Managed control-plane and session-data hosting from Private customer-hosted placement. Review session handling separately from inference.

### What is different about airgapped operation?

It adds a no-runtime-connectivity requirement. The Factory airgap build uses customer-configured models and disables Factory-bound services, and features such as Slack integration and hosted analytics are unavailable.

## Further reading

- [On-prem AI development](https://factory.com/articles/on-prem-ai-development)
- [Deployment patterns](https://docs.factory.com/enterprise/network-and-deployment)
- [Privacy and data flows](https://docs.factory.com/enterprise/privacy-and-data-flows)