A hosted agent is a vendor's bot working inside that vendor's platform, using the integrations the vendor supports. An owned agent runs as its own service on a host you control, usually as its own Linux user, with a purpose-built toolset and credentials you hold. Hosted gets you something running this week. Owned gets you inside the process that actually causes the work.
Three questions separate them: what the agent can reach, whose keys it holds, and what happens on the day you cancel.
What can each one reach?
A hosted agent reaches the systems its platform has connectors for, with whatever permissions the platform offers. An owned agent reaches anything you can give it a scoped credential for, which includes the internal API, the legacy system and the database that no connector covers.
When the process crosses those systems, a hosted agent cannot finish the job without a person carrying it between the steps.
Who holds the credentials?
On a hosted agent the platform holds the tokens, and your agent acts as one row inside a shared tenancy. An owned agent keeps its credentials in its own home directory on your host, scoped to that agent and revocable on its own.
Contractual isolation and technical isolation are different things, and only the second one survives an incident at the vendor. The technical version is set up as described in what agent isolation actually means.
Where does the data go?
A hosted agent sends your inputs to the vendor's platform and to whichever model that vendor wraps. An owned agent sends them where you point it: a hosted model, an inexpensive open-weight model by API, or a model running on your own hardware.
That last option is what makes an owned agent usable for privileged or regulated work, where a hosted bot is often not an option at all.
What happens when you cancel?
Cancel a hosted agent and the automations stop, while the prompts, the integrations and the history stay behind. Cancel an owned build and the system keeps running, because you hold the files, the credentials and the documentation.
The test worth applying before you sign anything: on the day you stop paying, what still works?
When is a hosted agent the better choice?
When the work is generic, the data is not sensitive, the platform's connectors cover it, and you want it live this week. Those are real cases, and buying a server you do not need is worse than paying the platform.
Owning becomes the answer when the process is specific, crosses systems with no integration, holds privileged data, or has to survive the vendor's next pricing decision.
Which model does each one use?
A hosted agent arrives with its model attached. An owned agent is model-agnostic: an open-weight model over an API where the volume is high, a frontier model where the step needs judgement, and either one changed by configuration without a rebuild. Open-weight versus frontier models covers how we pick between them.
Name one process that still needs a person. We will tell you what a one-time build takes.
Contact us today