Spike an AgentCore deployment and catch five wrong architecture assumptions

Earl Grey, a court operations manager who does not type code, had Kiro build and Claude review a throwaway deploy of a civic agenda watcher; the spike found five wrong assumptions and a redaction control that never ran.

Done with

Custom agent

Apps touched

AWS

ReportedSep 1, 2026

What they did

The poster, who runs court operations for a California county and builds civic tools without typing code, is making Porch Light, an agent that watches one city's public meeting agendas. Kiro writes the code, Claude reviews it against the poster's steering files, and he approves. Block Zero was meant to be a two-hour spike to prove the stack: invoke a model, run a Strands agent, deploy it to Bedrock AgentCore, confirm the logs land in CloudWatch. The folder was tagged throwaway. Block One then built the parser for Ventura's agendas, after a ten-minute check of the city's robots.txt and terms of use.

How it went

The spike took a build day. Kiro stopped and reported five places where the architecture document was wrong about AgentCore, from the deploy mechanism to the Python version to log group names, each cheap to fix then and expensive later. The deploy also showed that the log redaction module, with 28 passing tests, had never run inside the runtime because the spike used the framework's own logger. Both agents over-built the throwaway with a byte-identity test and dependency pinning, and he approved it. A scrub of his AWS account ID from git history kept failing because the number was quoted in the prose describing the scrub. Block One passed 57 tests, but a hand check found every meeting's documents listed twice.

Worth knowing

Kiro is AWS's agentic IDE and is not in our agent list, so this entry is tagged custom. The author's fix was a written rigor budget in the steering files: spike, working and full tiers, each a list of prohibitions the agent can be held to. He also found that the Granicus host behind the city's agendas disallows all crawling in robots.txt, which killed the only structured API.

Try it yourself withTry it withCustom agent

6 parts to fill in

This is a throwaway spike, tier: spike. Goal: prove that [the stack, e.g. a Strands agent deployed to Bedrock AgentCore] works end to end and that logs land in [CloudWatch]. Rules: no test files, no dependency pinning, no sync scripts, no refactors outside [this folder]. Read [the architecture document] first and, wherever the real platform disagrees with it, stop and report the discrepancy instead of adapting quietly. After the deploy, show me the actual log output and confirm whether [the logging or redaction module] really executed in the runtime. Budget [two hours]; stop and report as soon as the yes or no answer is known.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.