The agent is handling: "Cancel order #A-7821 and email the customer a refund confirmation." Click through each step to watch the agent reason, call tools, and work toward the goal.
Step 0 / 9
Click "Start" to begin
Ready to start
Click Start below or press → to walk through how the agent handles the task. You'll see reasoning steps, tool calls (ACT), and tool results (OBSERVE) as they happen. The dashed red border around the agent indicates that Guardrails wrap every input and output.
Step 0 / 9
Why this matters for the exam
The loop IS the value
A single FM call answers a question. An agent keeps going until the goal is achieved. Each Reason step decides what to do next based on observations. The FM never writes to your systems directly — it calls tools (action groups) that you control. This separation is why agents are safe enough to trust with real actions.
Order of operations
Notice the agent chose to look up before modifying. It didn't just cancel the order — it verified it first. You don't program this ordering; the FM reasoned it. This is exactly why agents can handle edge cases that scripted automation can't: when the order doesn't exist, the agent responds gracefully instead of failing a rigid script.
Exam angle
When a question describes "multi-step task,""plan and execute,""take autonomous actions," or "tool use" — the answer involves Bedrock Agents with the ReAct pattern. Expect action groups (Lambda-backed), OpenAPI schemas, and Guardrails in the correct option. See Pattern 3 for the full architecture.
Bedrock Agent Tracing
Every Reason / Act / Observe step in this loop is captured as a trace event in CloudWatch Logs. That's your debugging and audit trail — critical when an agent does something unexpected. On the exam, if a question asks about debugging agent behavior or explaining agent decisions, the answer involves Agent Tracing.