Most AI agent deployments put a frontier model in the request path. The result is high latency, variable answers, and no durable audit trail. But the knowledge the agent needs — refund policies, ticket history, escalation rules — already exists inside the business. It is just illegible to machines.
The better split is to do the messy knowledge work at build time: ingest docs, tickets, workflows, and code into structured skills. Then compile the repeatable decisions into a deterministic runtime with explicit predicates, no live model in the decision path, and every transition logged.
Interactive memo
The original positioning memo below includes the benchmark, the argument, and a side-by-side demo.
Compile once, run many times. The same input produces the same answer by construction, uncertain cases route to a human with full context, and policy changes ship like code — reviewed and versioned, not prompt-tweaked and prayed over.