Six weeks suits most IT teams. IT administrators need around 18 hours of learning, two weeks of supervised decisions on real access, connection and pause requests, and a final exam and sign-off. Their decisions are fewer but reach wider: one wrong connection can give an agent sources, or a role scope, it was never meant to have.
What an IT administrator supervises
IT administrators rarely judge an agent's draft letter or refund. They manage what sits around the agent: who has access to the workspace, which systems the agent connects to, which accounts it acts through, and how fast it can be paused. Those settings feed two of the five bounds every action meets, role scope and sources.
That changes what they need to learn. An administrator must understand that connecting a new system is not neutral. A ledger connection gives the agent sealed facts; an inbox connection gives it external ones, and an external claim cannot authorise a payment. They also need to know which actions the platform withholds for every agent, so they do not spend effort building blocks that already exist.
How the six weeks divide
- Week 1: Foundations, about four hours. Depth, the decision chain, provenance, ownership, data and trust.
- Weeks 1 to 4: Operator learning, about 14 hours. Weighted toward reading the workspace, when something goes wrong, and usage and billing.
- Weeks 4 to 6: Supervised decisions on real requests: access changes, new connections, pause and resume events, and certificate alerts.
- Week 6: Exam and sign-off.
Because requests are fewer, administrators should also run at least one pause drill on a non-production agent during the supervised weeks.
What the agent may do and what stays with IT
| Area | The agent | Stays with the IT administrator |
|---|---|---|
| Service desk tickets | Reads and drafts replies, resets within policy (depth 2 to 3, where earned) | Privileged access, new accounts, exceptions |
| Access reviews | Drafts a list of stale accounts for review (depth 2) | Removing access from a live user |
| Agent connections | None; agents do not connect themselves | Adding or removing any connection |
| Pausing an agent | None | Pausing on request from an owner or in an incident |
| Usage alerts | Reads usage and drafts a summary (depth 1 to 2) | Deciding whether to throttle or raise limits |
The rule that matters most: an agent never widens its own scope or adds its own sources. If a request to do so arrives, it is an incident.
An illustrative rollout
Take a mid-sized firm with an IT team of five and two agents in finance and customer service. All five finish Foundations in week one. Operator learning takes three weeks at about five hours a week.
During supervised weeks, each administrator decides real requests with the IT manager checking. One request asks to connect the finance agent to a shared mailbox so it can "see supplier updates". The administrator, now trained, recognises that every message in that mailbox would arrive as external provenance, and asks the finance owner whether the agent's brief expects that. The owner withdraws the request. In week six, the team runs a pause drill and sits the exam.
Calibration among administrators is covered in calibration sessions for IT administrators. A week-by-week plan is in the training plan template for IT administrators.
How this compares with other roles
Payroll administrators share the administrative mindset but review far more individual decisions; see how long it takes to train payroll administrators. Collections teams have high volume and fast feedback, so their timelines are shorter; see the collections agent timeline. The full approach is in training teams for AI agents.
Where the Academy fits
IT administrators who manage a live workspace day to day fit the Operator track of Pryme Intelligence Academy, after Foundations. Its labs run on the real product in a free Build workspace, so an administrator can practise pausing an agent and reading the resulting record without risk to production.
Questions people ask
Should IT administrators train before business reviewers?
At the same time, at least. IT sets the connections and access that reviewers depend on, and a reviewer cannot understand a held action if its sources were set up wrong.
Who should be allowed to pause an agent?
The agent's named owner, a deputy, and the on-call IT administrator. Make sure all three know how, and that the record shows who paused and why.
What should IT do when an agent's certificate lapses?
Nothing to the agent itself: a run that clears no depth drops it to depth 1 automatically. IT should make sure the owner knows and that held work is routed to people who can clear it.