Choose a model, give it a defined set of tools, and make its execution loop observable. A runtime turns those parts into an agent that can work inside an application, command line, or background process. The useful boundary is a session with inputs, events, artifacts, and a reason for stopping. That makes the worker replaceable and the surrounding factory reusable.
What this makes possible
- Embeddable sessions with typed events and explicit completion states
- Interchangeable model providers, tool sets, and execution loops
- Isolated workspaces for concurrent attempts
- Scripted model turns that exercise real tools in regression tests
On the production floor
A maintenance factory gives one agent a failing test and a disposable checkout. The agent returns a patch, a run log, and the result of the required checks.
How you know it works
Exercise the actual tool loop with known inputs. Check the resulting files, stopping behavior, and failures before introducing a live model.