All walkthroughs / Health organisation
An AI agent that ranks radiology documents is given a new version of its code and run. The run waits until the custodian of the Radiology department grants it access, in the same way a person's request would wait, and a second run has to ask again.
An agent is a program registered on Munitas, with a named owner, that Munitas runs on somebody's behalf. Registering an agent gives it no right to read data. When it needs a Raw dataset, its run waits until the custodian of the department that owns the dataset decides.
radiology-intake-triage ranks radiology documents so that the most urgent ones reach a human reviewer first. It belongs to Radiology, so the custodian of Radiology decides what it may read, and no other department's custodian does.
An agent's code is kept as sealed versions, the same way a dataset is. When it runs, the agent acts under an identity of its own, shown here as its runtime identity.
Devi chooses a zip of a small project. The model field says none, because the code calls no AI model, and the only tool it may use is read_dataset_version. The list of tools is a declaration. What limits the agent is that every read it makes is still checked against the same rules as a person's.
read_dataset_version.The version is identified by a fingerprint of its own code, which changes if any part of the code changes, and it is sealed on arrival. The platform runs it in a container, an isolated environment with no network access beyond its own requests for access.
After deploying the new version, Devi names a purpose for a run and chooses radiology-reports, a Raw dataset that Radiology owns. Munitas warns before the click, not after.
The run is recorded and goes nowhere until the custodian decides. That is not a failure. It is the same pause that a person's request would meet.
Okonjo is the custodian of the Radiology department. The request names both people involved: Devi asked, on behalf of the agent's own runtime identity, to read the Radiology dataset. Nothing from Cardiology appears in this list.
Okonjo's grant covered one purpose and was not a standing credential. Devi starts a second run on the same dataset for another purpose, and it waits again. This is what Raw data always gets. Reviewed data, which is no longer Raw, can be granted more widely.
Okonjo grants the second request, which this page does not show as its own screen, and the waiting run resumes by itself. Under each green label is the text that the agent's code printed. The sample agent prints one fixed result.
contains an unusually urgent phrase, flagged for a human reviewer first.