3.1.4. Agents for Amazon Bedrock — Multi-Step Task Automation
First Principle: A foundation model that can only generate text is passive — it answers questions. An agent transforms the same model into an active participant that can plan, decide, and execute sequences of actions to complete a goal.
Think of the difference between asking a colleague "What's the weather in Paris?" (passive retrieval) versus asking them to "Plan and book the best travel itinerary for my Paris trip next week" (multi-step planning and execution across multiple systems). Agents unlock the second capability.
How AI Agents Work:
- Goal received: The user provides a high-level goal in natural language (e.g., "Find the top three open support tickets, summarise each, and draft email responses").
- Planning: The agent uses a foundation model to decompose the goal into a sequence of smaller, actionable steps.
- Action execution: For each step, the agent calls the appropriate tool — this could be an API call, a database query, a Lambda function, or a knowledge base lookup.
- Result synthesis: After executing all steps, the agent synthesises the outputs into a coherent final response.
Why it matters for the exam:
The exam explicitly tests the distinction between a plain FM call (single-turn, text in → text out) and an agent (multi-turn, goal in → actions → result out). Agents introduce two key concepts:
- Tools: Definitions of what APIs or functions the agent is allowed to call. In Amazon Bedrock AgentCore, AgentCore Gateway turns existing APIs and Lambda functions into MCP-compatible tools.
- Knowledge bases: Data sources the agent can query for grounding (often backed by the vector databases covered in 3.1.3).
Contrast with RAG:
| RAG | Agents | |
|---|---|---|
| Input | User question | User goal |
| Output | Grounded answer | Completed task |
| Model role | Synthesise from retrieved context | Plan and orchestrate actions |
| Execution | Single retrieval + generation | Multi-step planning + tool calls |
Business Applications of AI Agents
Agents earn their cost when a goal needs several steps across several systems. Typical business applications include:
- Customer service: resolving an issue end to end, such as checking an order, arranging a replacement, and confirming by email.
- IT operations: investigating an incident, gathering logs, and proposing or applying a fix.
- Software and modernization: planning and changing code (Kiro), or migrating and modernizing workloads (AWS Transform).
- Research and analysis: gathering information from several sources into a cited report (Amazon Quick Research).
When a task is a single step, such as summarizing a pasted transcript, translating a paragraph, or classifying sentiment, a direct model call is cheaper, faster, and easier to govern than an agent.
Service status note: This section's title uses the original service name. Amazon Bedrock Agents is now Amazon Bedrock Agents Classic and has been closed to new customers since July 30, 2026; AWS recommends Amazon Bedrock AgentCore for new agents (see 2.3.3). The concepts in this section, such as planning, tools, and knowledge bases, apply to agents on any platform.
Scenario: A user asks an AI agent to "Pull last month's sales report from the database, identify the three lowest-performing regions, and draft a summary email for the sales director." The agent: (1) calls the database API to retrieve the report, (2) analyses the data to find the lowest performers, (3) generates a draft email. Without agents, this would require a human to orchestrate three separate steps.
Reflection Question: When would you use an AI agent instead of a simple RAG pipeline? What is the key signal in the requirements that tells you an agent is needed rather than a single retrieval + generation call?
⚠️ Exam Tip: The exam tests when agents are appropriate: look for requirements involving multiple steps, taking actions on external systems, or completing goals that require orchestration — not just answering questions from a knowledge base.
