The Terraform failure that costs me a day is never a syntax error. It passes terraform validate, tflint, and the HashiCorp MCP and is still wrong, and nothing disagrees.

So I make several models disagree, with an arbiter that committed first. Deliberation runs in my agent and has it write a verdict in the transcript before any reviewer sees the plan, then sends the plan to GPT, Gemini, Grok, and any model I added, in isolated threads under different reviewer roles. It triages each objection and loops until every reviewer and the verdict agree, or reports they did not.

The demo: no public access behind CloudFront with a WAF. The plan does that; two reviewers sign off. The Security Analyst refuses, HIGH: the execute-api endpoint is reachable on its own, so the WAF is decorative without a secret header binding CloudFront to the origin. Same shape on unenforceable IAM, a missing sts:ExternalId, and an S3 Deny that drops CloudFront. I also show how I measure which model earns its slot.

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.