Begin with a difference map, not a course. The team already knows how verdicts, depth, provenance and reasons work from the first agent; that transfers. What does not transfer is the new agent's capabilities, depths, sources, limits and red lines. Teach those differences, calibrate on cases where the two agents behave differently, and re-sign off reviewers for the new work.
What transfers and what does not?
A team that has run one governed agent for months has learned the hard part. They read verdicts and the bound that decided them. They know that held means a named person decides, that withheld is a platform decision nobody can release, and that depth is earned on evidence. They write reasons. They know the daily rhythm of the workspace.
The new agent changes the specifics. It may act at a different depth for similar-looking tasks. Its sources may be stronger or weaker. The business's limits for it will differ. Its red lines are new. Its certificate will lapse on a different date, so the team now tracks two expiry cycles. And where agents pass work to each other, reviewers must know which agent took which step.
The team lead owns the programme. The builder of the new agent provides the difference map. The head of the function signs off.
Why is habit the main risk?
People generalise from what they know. If the first agent issues refunds at depth 3 without review, reviewers will assume the new agent's payments work the same way, even if it only drafts them at depth 2. Or they will expect the new agent to hold anything over a limit because the first one did, and miss that its limit is set differently.
The result is not ignorance but confident error. Training for a second agent must therefore focus on contrasts: same-looking situations where the two agents behave differently, and why.
The steps
- Person (builder): writes the difference map: for each capability of the new agent, its depth, sources, the business's limits, red lines, and the nearest equivalent on the first agent if one exists.
- Person (team lead): marks each difference as "teach", "practise" or "red line".
- Person (team): completes a short session on the differences only, not a repeat of foundations.
- Person (team lead): runs contrast calibration on 30 paired cases, each pair showing the two agents handling a similar request differently.
- Agent, depth 1: runs the new agent read-only on live work for several days, so reviewers see its proposed actions alongside the first agent's real ones.
- Agent, depth 2: drafts live work for the new capabilities; reviewers decide in shadow while a reviewer already signed off on the new agent decides for real.
- Person (team lead): debriefs mismatches, labelling each as a habit error or a new-skill error.
- Person (head of function): re-signs off each reviewer for the new agent's decision types and updates approval rights.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Read-only period | New agent proposes; first agent acts as before | 1 | Team compares |
| New capability, drafts | Drafts | 2 | Reviewer sends or declines |
| New capability, held items | Prepares and holds | held | Reviewer signed off on the new agent decides |
| Handoff between agents | Each agent records its own step | 1 | Reviewer reads which agent did what |
| Overlapping request | The agent whose scope covers it acts; the other refuses | refused | Reviewer routes if needed |
How do you build contrast cases?
Pair requests that look alike but cross a boundary. For example, a refund request handled by the support agent at depth 3, and a supplier credit note handled by the new finance agent, which holds every credit for approval. Or a customer address change accepted by the first agent from the account portal, and the same change refused by the new agent because it arrived by email and the agent marks that source external.
Ask reviewers to predict each agent's verdict before revealing it. Wrong predictions show exactly where habit will bite.
Which controls matter, and what evidence do they leave?
- Difference map. Evidence: the signed map, versioned with the new agent's brief.
- Separate approval rights per agent. Evidence: the admin view showing rights for each agent and each capability.
- Re-sign-off. Evidence: the reviewer's results on the new decision types.
- Two certificate calendars. Evidence: expiry dates for each agent's certificates, reviewed weekly.
- Append-only record. Evidence: each step attributed to the agent that took it.
What should you measure?
Prediction accuracy in contrast calibration. Habit errors versus new-skill errors in shadow review. Red-line misses on the new agent, which must be zero. Time to decide the new agent's held items. And any drop in quality on the first agent's work while the team learns the second.
Where does it usually go wrong?
The team is sent back through the full introductory course. They are bored and learn nothing about the differences.
No one owns the difference map, so the team learns the new agent's limits from incidents.
The first agent's quality slips while attention is on the second. Keep measuring both.
New joiners are a related problem in reverse; see training new joiners when agents already run part of the work and onboarding a new support team lead to an agent workspace. For readiness and timing, see what go-live readiness is, what a team should practise before go-live and scheduling training around a go-live. The guide to training teams for AI agents puts it in context.
Where the Academy fits
Pryme Intelligence Academy teaches the concepts that transfer once: the Foundations and Operator tracks cover the decision chain, depth, held and withheld actions and reading the workspace. Module tests are set in the learner's industry and role. Teams on paid plans see each reviewer's results and certificate expiry in the admin view, which helps when two agents are live.
Questions people ask
Does the team need a new certificate for a second agent?
A person's certificate covers the concepts and stays valid for its period. The business should still re-sign off reviewers on the new agent's decision types before granting approval rights for them.
Should the same reviewers cover both agents?
Often yes, which is efficient but raises habit risk. Some teams assign the new agent's held items to a smaller group first, then widen.
What if the two agents disagree on the same case?
Each acts only within its own scope, and the record shows which acted. If both could act, the scopes overlap and the builder should fix that.