You add an agent by giving it a role, connecting only the sources and tools that role needs, certifying it on your own past cases, and starting it at a low depth that rises only with evidence. Each step has a named person responsible and leaves a record. Skipping certification and configuring a depth directly is the mistake to avoid.
What are the steps, and who acts?
| Step | Who acts | Evidence it leaves |
|---|---|---|
| 1. Define the role | Business owner of the process | Role record: purpose, capabilities, what is out of scope |
| 2. Assemble the brief | Business owner, with the platform | Brief version: instructions, sources, connections |
| 3. Register sources and tools | Platform team, approved by system owners | Connection records, each tool mapped to a capability |
| 4. Set your own limits | Risk or control owner | Limit set per capability: amounts, counts, times |
| 5. Certify on your cases | Platform runs it; owner reviews results | Certification report: score per gate |
| 6. Approve for production | Named approver, not the builder | Approval record with the depth granted |
| 7. Start at the earned depth | Platform enforces | Every verdict filed from the first action |
| 8. Monitor and renew | Owner and platform | Drift results; recertification before expiry |
What does each step involve?
1. Define the role. Write what the agent is for in one or two sentences and list each capability separately: look up, draft, update, send, pay. A role that cannot be described briefly is probably two roles.
2. Assemble the brief. The brief turns the role into instructions the agent follows, the sources it may use and the systems it connects to. It is a candidate at this stage, not a deployed agent.
3. Register sources and tools. Each tool maps to one business action and one capability. Each source is stamped with how far it can be trusted. System owners approve every new connection. Agents that touch payments need extra care here; whether an AI agent can add a new payee walks through one such capability.
4. Set your limits. These are your own business rules: a maximum refund, a cap on changes per day, no actions outside office hours. They sit alongside the depth the agent earns, and the lower of the two applies.
5. Certify. Replay the agent against real cases from your own operation and score each answer. Certification is per agent: the agent may clear depth 2, enough to "draft reply", but not depth 4, which "issue refund" needs. The result is one depth, the highest it cleared.
6. Approve. A named person who did not build the agent reviews the certification report and approves production use at the earned depth. The role's ceiling still applies above that.
7. Start. From the first live action, every verdict is filed. Where a capability needs a person, add the gate as described in how to add a human gate to an agent workflow.
8. Monitor and renew. Certificates expire. Plan recertification before they do, and watch for drift in between.
What does this look like in practice?
Take a mid-sized lender adding an agent to handle direct debit date changes. The collections manager defines the role: read the account, check the requested date against the lender's rules, update the date, send a confirmation. Four capabilities.
The platform team registers two tools against the loan system and one against the messaging system. The risk owner sets a limit: no more than one date change per account in a rolling period. Certification replays several hundred past requests. The agent clears reading and drafting comfortably, clears updating, and is borderline on confirmations because it sometimes omits the first payment amount. The approver grants depth 2, the highest depth the run cleared, so a person approves updates and sends confirmations until the brief is fixed and the agent recertifies.
What stays with a person?
The role, the limits, approval for production, every new connection, and any increase in depth. Agents are added to the platform by people; they do not add themselves or each other.
How does this differ from adding an agent to an agent platform?
On many platforms, adding an agent means configuring it and switching it on, with permissions set by whoever configures it. In a governed operating system, permissions are earned through steps 5 and 6. The wider difference is set out in how an agentic operating system differs from an agent platform, and the agentic operating system requirements checklist helps test any option. The full design is in the governed agentic operating system guide.
Where Pryme fits
In Pryme, Agent Studio assembles the brief from a plain-language job description and produces a candidate. Attest certifies each agent by replaying it on your cases against seven gates. The agent's role sets a ceiling, the evidence sets the depth, and the lower wins. Certificates lapse after 90 days.
Questions people ask
How long does adding an agent take?
It depends mainly on connections and on having enough real cases to certify against. Defining the role and brief is usually quick; integration and assembling good test cases take the time.
Can we add an agent built by a vendor?
Yes. It goes through the same steps. A vendor's own testing does not replace certification on your cases.
What if the agent fails certification?
It stays at the highest depth it cleared, or does not go live. The report shows which gate failed, which tells you whether to change the brief, the sources or the scope.