Nothing moves without permission. Decide what each person and each agent can read, write, and run, per system, per project, per individual action.
Decide what every agent can do
Permissions are set per person, per agent, per system, and per individual action, and each one can be tuned without touching the rest.
User-level access
An agent acts as the person who asked, and it cannot reach anything that person's own account cannot reach.
Per-action permissions
Set every tool to always allowed, held for approval, or never available, one action at a time.
Per-system scopes
Decide for each connected system whether an agent can search, create, update, or execute there.
Require approval where it matters
Consequential steps in a validated system stop and wait for a named person before they run.
Approval gates
A write to a regulated system waits until the named reviewer approves it, in the same thread as the work.
No silent writes
Nothing lands in the QMS, the eTMF, or the ELN without the sign-off that system's own process requires.
Instant revocation
Remove a person's access and every agent acting as them loses it at the same moment, including mid-run.
Make every run reviewable
Every run keeps the person who asked, the instruction they gave, each action taken, and the approval that released it.
Full attribution
Every tool call stays tied to the account it ran as, the system it touched, and the run that produced it.
Source trail
A result keeps the documents, records, and data behind it, so it can be checked without redoing the work.
Run history
Open any run months later and see exactly what the agent did, step by step.
