A blog where I record notes and learnings about technology.
Shinya Tahara
Solutions Architect @ AWS
I'm a Solutions Architect at AWS, providing technical guidance primarily to financial industry customers. I share learnings about cloud architecture and AI/ML on this site.The views and opinions expressed on this site are my own and do not represent the official positions of my employer.
Create 4 types of payment cryptographic keys with Java SDK, test wrong-key-usage errors, and experience how TR-31 KeyUsage enforcement differs from KMS.
Measured real downtime during Aurora PostgreSQL failover vs Blue/Green Switchover. Failover had 12 connection failures in two separate windows, Blue/Green Switchover ~26s with 6 failures. DNS TTL of 60s is the key bottleneck.
A 5-level maturity model for coding agent adoption, built on Anthropic and OpenAI frameworks but tailored to software development teams. Assess your current level and identify the highest-impact next step.
Walk through the Strands Agents SDK Python Quickstart hands-on, explaining the agent loop, custom tools, and metrics with working code. A few dozen lines of Python is all it takes.
Tested Stateful MCP's 3 capabilities with a DynamoDB-backed expense tracker. Verified data integrity on Elicitation decline/cancel and fallback behavior on Sampling failure to provide production design guidance.
Hands-on verification of S3 account regional namespaces via CLI. Measured per-region prefix length limits, tested IAM condition key enforcement, and discovered built-in name protection in the global namespace.
Verified DevOps Agent's EKS integration. Compared investigation processes for application vs. infrastructure failures, measured audit log-based operator identification, and evaluated preventive recommendations for critical system components.
Run the same task with Agents as Tools, Swarm, and Graph, then compare structural differences via metrics. Establish selection criteria for choosing the right pattern per use case.
Embed a Swarm as a node in a Graph to combine autonomous collaboration with structured workflows. Verify nested execution results and multi-agent Hooks for node monitoring.
Strands Agents SDK Graph with cycles (feedback loops) fails build() auto-detection with ValueError. Use set_entry_point to specify the starting node explicitly.
Strands Agents SDK Graph's result.execution_time returns 0ms. Per-node execution_time is correct, so calculate the total yourself.
In Strands Agents SDK Swarm, agents that hand off via handoff_to_agent have no text block in their result. The final text output comes from the last agent that didn't hand off.
AWS Agent Registry CLI uses skillMd (dict with inlineContent) while REST API uses skillMarkdown (string). descriptorType is AGENT_SKILLS, not SKILL.
AWS Agent Registry MCP records require reverse-DNS name format, 100-char description limit, and explicit schemaVersion. The error message doesn't tell you which field is wrong.