Agents
From answering to acting
The story: A hotel receptionist who can only tell you "yes, we have rooms" is helpful. One who can also book the room, charge your card, and print your key card is far more helpful.
In AI/AWS terms: Agents extend a generative AI model so it can act, not just answer.
For the exam: Agents let a model take actions, not only produce text.
Key functions
The story: The receptionist connects you to the hotel's booking system and billing system. They make changes: move your room, take payment, print your invoice. And they note what went well and badly, so the hotel improves.
In AI/AWS terms:
- Intermediary operations: connect the model to backend systems such as databases, CRM platforms, and service management tools.
- Launching actions: adjust settings, process transactions, retrieve documents.
- Feedback integration: collect the outcomes of actions to improve the model over time.
For the exam: Agent functions: connect to backend systems, launch actions, and feed outcomes back.
AnyCompany's three agents
The story: The phone company gives its chat assistant three helpers: one who changes customers' plans, one who files notes from every conversation so the binder keeps growing, and one who hands out a "how did we do?" card at the end.
In AI/AWS terms:
- Agent 1 changes customers' plans based on their prompts.
- Agent 2 adds data from conversations to the enterprise database, which improves future RAG answers.
- Agent 3 sends a satisfaction survey when the conversation ends, to track the 4-out-of-5 goal.
For the exam: Agents can also feed RAG: saving conversation data improves later answers.
Multi-step tasks
The story: "Work out my total bill, then send it to me as a neat invoice" is two steps in a row, and the second depends on the first.
In AI/AWS terms: A multi-step task chains actions together, such as calculating a value and then formatting the answer as JSON.
For the exam: A multi-step task chains actions, where each step uses the result of the one before.