The half of the build nobody demos.

Infrastructure is where a working agent is separated from an impressive one. Integrations that write live during a call, conflict rules that stop double-booking, a fallback chain for provider outages, structured logging, and the evaluation suite that gates every change to any of it.

What this pillar owns

• Authentication, token rotation, and field mapping into your real pipeline.
• Conflict rules across multiple calendars and technicians.
• Model fallback with per-request health checks.
• Transcript, summary, disposition, and latency trace on every conversation.
• The regression suite that blocks a bad deploy.

Where builds actually fail

Not model quality. They fail because an integration silently stopped syncing for eleven days, or because two calendars disagreed and the agent booked a technician who was already on a job. Both are boring, both are expensive, and both are what most of the build fee is actually paying for.

What you can check yourself

Ask any vendor to show you a deploy their eval suite blocked. If no such deploy exists, either the suite is not run or it does not gate anything. That single question separates production engineering from a prompt in a console faster than any technical interview.