Copyright (c) 2026 MindMesh Academy. All rights reserved. This content is proprietary and may not be reproduced or distributed without permission.
4.4. Reflection Checkpoint
Key Takeaways
- Bedrock Agents manages the full reasoning loop; Strands Agents provides code-first customization; Agent Squad enables multi-agent coordination. Use agents only when tasks genuinely require multi-step tool use — simple Q&A over documents doesn't need agent overhead.
- MCP is an open protocol (not AWS-proprietary) that standardizes tool integration across agent frameworks. Without MCP: N clients × M tools = N×M custom integrations. With MCP: N clients + M servers = N+M implementations.
- Human-in-the-loop uses Step Functions
waitForTaskToken— the workflow pauses until a human provides approval viaSendTaskSuccessorSendTaskFailure. Standard Workflows incur cost while waiting; prefer task tokens with reasonable heartbeat timeouts. - API Gateway has a 29-second timeout — long streaming FM responses require Lambda Function URLs (streaming mode) or WebSocket APIs to avoid Gateway timeouts.
- Exponential backoff must have a maximum retry limit + fallback model. Unlimited backoff under sustained throttling cascades into Lambda concurrency exhaustion.
- Model routing cuts costs 60–90% by directing simple queries to cheaper models. Keep routing classifiers lightweight — FM-based classifiers double cost for simple queries.
Connecting Forward
Phase 5 covers Domain 3 — the security, safety, and governance layer that every production GenAI application requires. This domain's concepts pervade all other domains: Guardrails appears in Domain 1 prompts, Domain 2 agents, and Domain 4 monitoring. Governance principles appear in exam scenarios across every domain.
Self-Check Questions
- A Bedrock Agent is making 25 tool calls to answer a simple question about office hours. It's in a reasoning loop where it keeps calling the same knowledge base search tool with slightly different queries. What three controls would you add to prevent this?
- Your team ships a prompt update on Friday afternoon. By Saturday morning, production logs show the FM is now producing responses that include competitor product names — something that never happened before. Walk through the complete incident response: what you check first, how you roll back, and what you add to the CI/CD pipeline to prevent recurrence.
Written byAlvin Varughese
Founder•15 professional certifications