Ask Your AI Agent to Break Your Cloud App

An experiment used an AI agent with LocalStack’s MCP server, Chaos API, and App Inspector tracing to autonomously run a chaos-engineering game day against a Go/CDK serverless pipeline consisting of API Gateway, Lambda, Kinesis, DynamoDB, and CloudWatch. After establishing that 60 normal requests were delivered successfully, the agent injected Kinesis and DynamoDB outages, DynamoDB throttling, network latency, and CloudWatch failures, identifying that short Lambda timeouts produced ambiguous 502 errors, unlimited event-source retries could stall processing, and downstream metric-publishing failures were silently logged rather than reported. It updated the application to constrain SDK retries within Lambda deadlines, return clear 503 responses, process failed batch records selectively, return downstream errors, configure bounded retry policies, split failing batches, and send exhausted failures to a dead-letter queue. Repeated tests showed that the pipeline retained normal delivery, accepted all requests under 800 ms added latency instead of none, preserved records during DynamoDB outages through observable retries, and surfaced CloudWatch failures through invocation errors and DLQ entries. The unattended session took 26 minutes, used $8.23 in tokens, made 37 MCP tool calls, and demonstrated how compatible AI agents can reduce the cost and effort of cloud resilience testing without a production-like AWS staging environment.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.