4. Specialized Agents Workshop
At a Glance
| Format | Hands-on |
| Participants create | A specialized operational agent |
| Key message | The most valuable agents understand how your organization operates |
| Duration | to be defined |
Objective
Participants create an agent that brings organizational knowledge into DataMiner: company policies, operational procedures, escalation rules, standards, and best practices. The purpose is not AI experimentation.
What Is Given
| Item | Details |
|---|---|
| Same backend package | The vertical from the App Builder Workshop, with its data model and live scenario |
| Knowledge pack per vertical | Sample policies, procedures, escalation rules, and standards that match the vertical's data (for example the escalation thresholds in the data model) |
| Agent building environment | Access to the tooling used to build the agent – see Participant Enablement |
| Agent design card | Handout: role, scope, knowledge, allowed actions, guardrails, test questions |
All AI output lives in this session – for example service summaries, recommended actions, and an Operations Copilot. It is generated at runtime and not stored in the backend.
Before the Session
- [ ] Knowledge packs available for all five verticals
- [ ] Agent tooling access tested for every participant
- [ ] Backend scenario still running (or restarted) so agents see live incidents
- [ ] A reference agent per vertical to demonstrate or unblock participants
Session Flow
| # | Trainer | Participants |
|---|---|---|
| 1 | Recap the app from the previous session and ask: "What does an operator still need to know that the data does not tell them?" | Identify missing organizational knowledge |
| 2 | Introduce the knowledge pack of each vertical | Read the policies and procedures for their vertical |
| 3 | Explain the agent design card | Define the agent's role, scope, and the question it answers |
| 4 | Stress guardrails: what the agent may and may not do | Define allowed actions and human approval points |
| 5 | Support building | Connect the agent to the data model, real-time data, and the knowledge pack |
| 6 | Provide tricky test questions | Test the agent, e.g. "What requires escalation according to our policy?" |
| 7 | Run a short show & tell | Show how the agent applies the organization's rules |
Example Operations Copilot questions:
- What requires immediate attention?
- Which service is at highest risk?
- Summarize the operational situation.
- What should I investigate first?
Do's and Don'ts
| Do | Don't |
|---|---|
| Ground every answer in the knowledge pack and live data | Build a generic chatbot |
| Keep the agent's scope narrow and clear | Try to make one agent do everything |
| Test with questions where policy changes the answer | Only test easy questions |
| Require human approval for impactful actions | Let agents take uncontrolled actions |
| Treat AI output as runtime output | Store AI output in the backend |
Success Criteria
- Each agent answers at least one operational question according to the organization's rules, not just based on data
- Participants can explain what their agent may and may not do
Transition
"One agent that understands your organization is powerful. Next, we'll show what happens when agents with different responsibilities work together – with humans in control."