Topic 7
Workflow or agent? And wrap-up
Here's the same choice on ten illustrative support tickets. The workflow has three fixed paths (duplicate charge, cancellation, change card); it classifies each ticket and either runs a path or hands the ticket to a person. The agent chooses its own steps for every ticket.
Predict first
The same ten tickets, two designs
Illustrative tickets and outcomes · tallies computed
Resolved
5 / 10
Handed to a person
5
Wrong action taken
0
Model calls
15
| Ticket | Outcome | Model calls |
|---|---|---|
| Charged twice for order 4417 | resolved | 2 |
| How do I cancel my plan? | resolved | 2 |
| Refund for an order from last week | resolved | 2 |
| Charged twice, and also want to cancel | handed off | 1 |
| Change the card on my account | resolved | 2 |
| Refund one item from a three-item order | handed off | 1 |
| Was I charged tax correctly? | handed off | 1 |
| Cancel, but keep my files | handed off | 1 |
| Why is my invoice in euros? | handed off | 1 |
| Duplicate charge on a gift purchase | resolved | 2 |
Neither row is “the answer”. If a person is cheap and a wrong action is expensive, the workflow wins. If most tickets are unusual and handoffs are slow, the agent may be worth its cost, but only with the harness of the coming lessons around it. A common middle path: run the workflow for known cases and send only the rest to an agent, so the expensive, error-prone path handles only what needs it.
Wrap-up. From this lesson, you can now:
- tell a workflow from an agent by who chooses the next step;
- use the five workflow patterns before reaching for an agent;
- decouple planning from execution, and validate plans in code first;
- name the planning failures a plan check catches, and the one it can't;
- describe sequential, parallel, conditional and looping control flows, and why model-judged loops need limits;
- set a level of automation per action, with people at the risky ones;
- compare designs on resolution, handoffs, wrong actions and cost.
Quick check