Build a small central team that runs the machinery, and leave ownership of each agent with the business team whose work it does. The centre writes standards, certifies, monitors and keeps the register. Business teams own their agents, approve held actions and decide depth. Risk challenges both. Keep certifying apart from building, so no one marks their own work.
What does an agent operations team actually do?
Once more than one or two agents are live, someone has to run them as a service. That means keeping a register of every agent and its capabilities, running certification and renewal, watching performance, handling incidents, maintaining the briefs and sources, and making sure held work reaches a named person. None of that is the same as building agents, and little of it belongs to IT.
The decisions that stay with named people are few but important: which processes get agents, what depth each agent holds, when to widen or narrow, when to retire an agent, and how to respond when one fails. The team design should make clear who holds each of these.
What roles does the team need?
Head of agent operations. Accountable for the service: the register, the standards and the operating rhythm. Usually sits in operations, not technology, for the reason argued throughout our guide to deploying an AI agent: the people who own the work should own the agents doing it.
Agent owners, in the business. One named owner per agent, normally the manager of the process. They decide depth, approve changes to the brief, and answer for the agent's work.
Approvers, in the business. The people who work held queues. They are part of their own teams, not the central team, because they need to know the work.
Certification lead. Runs the replay of agents against real cases, sets the floors with each owner, and signs the result. Must not be the person who built the agent.
Agent analysts. Write and maintain briefs, curate sources, investigate failures and prepare the evidence for widening. This is the role most organisations underestimate. The skills it needs are set out in what skills a team needs to run AI agents.
Platform engineer. Keeps connections, service accounts and integrations working. Often shared with IT.
Risk partner. From the second line. Reviews the standards, attends depth decisions for higher-risk agents, and challenges.
How should it be set up, step by step?
- Executive sponsor: appoints the head of agent operations and states, in writing, that business owners own their agents.
- Head of agent operations: sets up the register. Every agent, its owner, its capabilities, its certified depth, certificate dates and approvers.
- Head of agent operations and risk partner: write the standards: how depth is set, how certification works, what is recorded, how incidents are handled.
- Each business owner: names approvers for their agent's held work, with cover.
- Certification lead: certifies existing agents against the standard before new ones are added.
- Head of agent operations: starts the weekly review, one meeting covering every live agent's figures, held queues and incidents.
- Agent, depth 1: produces the review pack from its own records and those of other agents, for the team to discuss.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Appoint lead | Nothing | none | Sponsor decides |
| Register | Keeps entries current from records | 1 | Owning the register |
| Standards | Nothing | none | Writing and approving |
| Approvers | Nothing | none | Naming them |
| Certify | Replays cases | 1 | Signing results |
| Weekly review | Drafts the pack | 2 | Decisions |
How big should it be?
Smaller than people expect at the start. With one agent in one team, the roles can be part-time hats on existing people, as in running a pilot of an AI agent in one team. The head of agent operations, a certification lead and one analyst can support several agents. The workload grows with the number of agents and their renewal cycle, not with the volume each agent handles.
What grows faster is approver time, which sits in the business. Each new held capability needs someone who can review it. Plan that time when the agent is approved, not after the queue builds.
What records does the team keep?
- The agent register, current at all times.
- A decision log: every depth change, its evidence and who decided.
- Certification files per agent, with renewal dates.
- Weekly review notes, with actions and owners.
- Incident records for agent failures, linked to the depth changes that followed.
Together these let anyone reconstruct why any agent was doing any piece of work on any day.
Where does it usually go wrong?
The central team owns the agents. It feels efficient, but the business stops feeling responsible for the work and the central team ends up making operational decisions it is not placed to make. The relationship between agents and the people around them is covered in managing a team of AI agents and people.
Builders certify their own agents. Every team believes its agent is ready. Separate the roles, even in a small team.
Second line arrives late. Risk is invited after the first incident. Invite them to write the standards.
Expansion outpaces the team. A second business unit wants agents, and the central team is stretched. Each new team needs its own owner and approvers before go-live; see extending an AI agent to a second team.
The weekly review becomes a status meeting. If decisions are not made in it, it will be skipped. Every review should end with recorded decisions, even if the decision is "no change".
Where Pryme fits
In Pryme, the register, certificates, renewal dates and every verdict are held in one place, which reduces the central team's clerical work. Attest gives the certification lead a repeatable replay, and Workspace gives approvers a single held queue across agents.
Questions people ask
Should the agent operations team sit in IT or in operations?
In operations, with strong links to IT. Agents do operational work, and the people who answer for that work should run them.
Can one person be both agent owner and approver?
For low-risk capabilities, yes. For anything that moves money or affects customers significantly, keep a second approver so the owner is not reviewing work they set the rules for.
How often should the register be reviewed?
Monthly in full, with certificate expiries checked weekly. An agent whose certificate lapses should lose depth automatically, not wait for the review.