Automating the Experimentation Lifecycle with Kiro, AWS DevOps Agent, and LaunchDarkly
AWS describes a reference architecture where AWS DevOps Agent runs a scheduled custom agent that plans feature-flagged experiments, uses Kiro CLI in a custom MCP server to write code and open pull requests, and deploys through GitHub Actions and AWS Amplify. LaunchDarkly runs the experiments and Guarded Releases, which roll back flag state when error or latency guardrails are breached. The post says the architecture is a reference implementation and that a more turnkey version is planned.
What changes for platform teams
A platform team evaluating agent-driven delivery would need to scope MCP tools, restrict repos, branches and allowed tools, and manage credentials and state outside the container. The post notes that the scoping lives in server code rather than prompts, and that MCP tool allowlists are recommended for production.
Read the original at AWS DevOps Blog ↗
written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules