1.3.2. AI Agents and Agentic Loops
💡 First Principle: An AI agent uses an FM as its reasoning engine but augments it with the ability to call external tools — APIs, databases, code interpreters — and use the results to inform subsequent reasoning steps. The FM doesn't just generate text; it generates decisions about what to do next.
Without agents, FMs are passive: you give them input, they give you output, and nothing changes in the external world. Agents break that limitation. An agent can look up your current order status, book a meeting, write and execute code, or trigger a deployment pipeline — all through an FM reasoning about which tool to call and with what parameters.
The ReAct reasoning loop:
Amazon Bedrock Agents manages this loop automatically — it handles session state, tool selection, error recovery, and response synthesis. AWS Strands Agents (open-source, announced 2025) provides a code-first framework for building agents with greater customization. AWS Agent Squad enables multi-agent coordination where a supervisor agent delegates to specialist sub-agents.
When to use agents vs. RAG:
| Scenario | Pattern | Why |
|---|---|---|
| "Answer questions about our documentation" | RAG | Passive retrieval only; no action needed |
| "Book a meeting based on calendar availability" | Agent | Must take action in external system |
| "Summarize this uploaded PDF" | RAG (or simple FM call) | Single-step, no tool use |
| "Research competitors, draft a report, email it" | Agent (multi-step) | Multi-step task requiring multiple tool calls |
| "Answer questions and update CRM with customer sentiment" | Agent + RAG | Both retrieval and action required |
⚠️ Exam Trap: Agents add latency, complexity, and failure surface area. The exam tests whether you recognize when NOT to use agents — a simple Q&A use case implemented as an agent introduces unnecessary overhead. Choose the simplest architecture that meets requirements.
Reflection Question: A team proposes using a Bedrock Agent to handle a use case where users upload a document and ask questions about it. No external API calls are needed. Would you approve this architecture? What would you propose instead, and why?
