The agent-written Terraform failure that costs me a day is never a syntax error. It passes terraform validate, tflint, and the HashiCorp Terraform MCP and is still wrong. One model does not catch that; nothing disagrees with it.

So I make three disagree. `consensus` sends the same artifact to GPT, Gemini, and Claude in fresh independent threads under different reviewer roles. Where they split is the risky part. Claude triages every objection, revises, and loops until all three sign off or reports they did not.

The brief said no public access, everything behind CloudFront with a WAF. The plan did that, and two reviewers signed off - it looks fronted. The third, "Security Reviewer", refused: the execute-api endpoint is still reachable on its own, so the WAF is decorative without a secret header binding CloudFront to the origin. The loop keeps the disagreement. Then an unenforceable IAM control and an S3 Deny that drops CloudFront. Same loop on Python, TS, marketing, and troubleshooting.

Target Audience:
Mid-to-senior DevOps, platform, SRE, and AI-engineering builders who run real infrastructure changes, especially on AWS via Terraform, and have already used coding agents on work that mattered.