Bedrock Prompt Shortening: 6 optimizations, 0 changes, a fake +45%
I tested Bedrock's shortenPrompt on my agent's real prompt across three models: it returned the original six times and reported a +45% gain with no change.
AWS architecture tested in production, now focused on AI agents
I'm Gerardo Arroyo, an AWS Solutions Architect and AWS Golden Jacket. I publish measured experiments on AWS, from serverless and databases to Bedrock and AgentCore: what works, what breaks, and what the documentation leaves out.
Articles
2026
Temporal Policies in AgentCore: 98 Denied Deploys and a 409 That Never Came
I tested AgentCore Temporal Policies with 5,294 calls: an agent that parallelizes loses its deploy, denials burn your limit, and the 409 never arrives.
I Took the Wheel Off My RAG: Managed KB vs. Manual Chunking
Managed Knowledge Base ties my manual chunking setup, but AgenticRetrieveStream's default planner never decomposed a single query across 80 evaluated questions.
I Let AWS Rewrite My DevOps Agent's Prompt: 142 Sessions, +6.2%, and a p-value of 0.38
AgentCore Optimization closes the observe → evaluate → improve loop. I ran the full loop on my DevOps agent: recommendation, A/B test, and the gotchas.
Lambda MicroVMs vs AgentCore Runtime: When to Use Each for Production Agents
AWS launched Lambda MicroVMs: isolated Firecracker VMs, state up to 8h. Is it the same as AgentCore Runtime? A decision guide for CTOs and cloud architects.
AgentCore Payments: When Your Agent Has Its Own Wallet
AgentCore Payments gives your agent its own wallet. I ran the x402 flow on a real AWS account: code, spend limits, and 4 gotchas before the first on-chain tx.
AWS Agent Registry: a private catalog to stop agent sprawl
Hands-on guide to AWS Agent Registry, now generally available (GA): 4 IAM personas, 8 real records, 12 gotchas, and a Claude Code demo consuming the catalog over MCP.
Real Benchmark: 5 Chunking Strategies in Amazon Bedrock Knowledge Bases
A reproducible benchmark with 25 questions, Sonnet 4.6 as generator and Nova Pro as cross-family judge. Plus the 7 problems I hit deploying this with Terraform. April 2026.
Your Coding Agent Will Never Start From Scratch Again: Session Storage in AgentCore Runtime
Every time you restart a coding agent, it loses everything: installed dependencies, generated files, project state. AgentCore Runtime Session Storage solves this with an AWS-managed persistent filesystem. Full implementation with Strands Agents and real, tested code.
Hi, I'm Gerardo
AWS Solutions Architect, author of AgentCore in Production: The Operator’s Playbook, and independent consultant specialized in Amazon Bedrock and AgentCore. AWS Golden Jacket, Community Builder, and former AWS Ambassador. I help technical teams take GenAI architectures from prototype to production — without the theory that’s already in the official documentation, with the real problems that only show up when code runs in a real AWS account. This blog is the public version of my work. If what I write resonates and you’re building something similar, let’s talk.
Building something like this?
I help technical teams take GenAI architectures on AWS from prototype to production: architecture reviews, implementation sprints and advisory on Bedrock and AgentCore.