2.2.2. AI Agents: Autonomy, Tool Use, and Orchestration
💡 First Principle: A chatbot answers; an agent acts. An AI agent is given a goal, plans the steps to reach it, uses tools (business systems, APIs, data sources) to carry them out, observes the results, and adjusts until the goal is met or it decides to escalate. Because it acts in your systems on your behalf, an agent's value and its risk both scale with its autonomy and its permissions.
Core capabilities named in the exam guide:
- Autonomy: the agent decides which steps to take rather than following a fixed script. Autonomy is a dial, not a switch: from "suggest actions for approval" through "act within limits" to "act freely." Higher autonomy increases speed and the need for guardrails.
- Tool use: the agent calls external capabilities, such as looking up an order, issuing a replacement, updating a ticket, or searching a knowledge base. Standards such as the Model Context Protocol (MCP) give agents a common way to connect to tools and data sources.
- Agent-to-agent communication: specialized agents collaborate, for example a customer-service agent asking a billing agent to verify a charge. Protocols such as Agent2Agent (A2A) let agents built by different teams or vendors exchange tasks and results.
- Orchestration strategies: how work is coordinated across steps or agents. Common patterns include a supervisor agent that delegates to specialist agents, sequential handoffs (one agent's output feeds the next), and parallel execution of independent subtasks with results combined.
On AWS, Amazon Bedrock AgentCore provides managed building blocks for agents: a runtime to execute them securely, memory, a gateway that exposes tools (including MCP servers and APIs), identity and authorization so agents act with the right permissions, and observability to trace what an agent did and why. Amazon Quick also offers agent and automation features (Quick Flows, Quick Automate) aimed at business users. You do not need to configure any of these; you need to recognize when a scenario requires an agent and what controls it needs. (The older Amazon Bedrock Agents, now called Bedrock Agents Classic, is closed to new customers; AWS points new agent projects to AgentCore.)
| Solution type | What it does | Takes actions in systems? | Best for |
|---|---|---|---|
| Rules-based workflow | Executes predefined steps and branches | Yes, only scripted ones | Predictable, fixed processes |
| Conversational chatbot | Answers questions from its training or a script | No | FAQs, simple guidance |
| RAG assistant | Answers using retrieved company content, with citations | No | Policy and knowledge questions grounded in current documents |
| AI agent | Plans and executes multi-step tasks using tools, adapting to results | Yes, within granted permissions | Variable, multi-step tasks needing judgment and system actions |
Consider a retailer handling damaged-order complaints. Each case requires checking order status, verifying the damage claim, choosing between replacement and refund based on stock and policy, executing it, and escalating unusual cases. A chatbot can only explain the policy. A RAG assistant can quote the right policy section. A rules workflow breaks on the variety of cases. An agent integrated with order, inventory, and payment systems, with spending limits and escalation criteria, fits the task.
⚠️ Exam Trap: When a scenario requires the system to take actions across business systems, decide among next steps, and adapt to results, the answer is an AI agent, not a chatbot or a RAG assistant, both of which only produce text. Equally, when the task is only answering questions from documents, an agent adds cost and risk without benefit.
Reflection Question: An agent can issue customer refunds. What limits, permissions, and escalation criteria would you require before approving it for production, and which of these are business decisions rather than technical ones?
