Most AWS debugging time goes into the console.
Click into a log group. Pick a stream. Squint at plain text. Filter by string match. Tab over. Compare timestamps. Repeat.
Three years ago that was my whole life. Then I went through three levels of observability — and stopped opening the console entirely.
Level 1 — Log Hell
console.log("Processing order " + orderId + " for customer " + customerId) scattered across 15 log groups. No correlation
IDs. No log levels. When a customer reports a missing order, you grep.
It works on a Tuesday afternoon. It breaks at 2am.
Level 2 — The Three Pillars
Structured JSON logs (AWS Powertools), EMF metrics, X-Ray traces. Every log carries the same correlationId from API Gateway → Lambda → SQS → Lambda →
DynamoDB. Logs Insights actually works — one query, both log groups, all events for one customer.
This is what AWS docs recommend. Most teams stop here. It's the prerequisite for what comes next.
Level 3 — A Claude Code Agent
One markdown file in .claude/agents/ that knows your log groups, knows your query templates, and talks to CloudWatch via
the AWS CLI.
I type: "find any orders that timed out yesterday."
It picks the right log groups. Writes the Logs Insights query. Polls for results. Groups errors by type. Tells me which customers were
affected. No console. No query syntax.
The full agent file is in the
free GitHub gist. Copy it, swap your log group prefix and AWS profile, you're running in 20 minutes.
Watch the video for the full walkthrough — the e-commerce order-processing demo, real seeded timeouts, and why Level 2 is the
prerequisite that makes Level 3 work.
