The Agent Validator Pattern — the 3 deterministic checks that took a 90% duplicate-refund rate to zero across 392 live agent runs

The companion playbook to the video. We ran 392 live runs against four models people actually ship agents on — Claude Sonnet 5, Claude Haiku 4.5, DeepSeek v4, GLM 5.2. With rules visible in the prompt they were perfect (0/312, even under social engineering). When one field went missing from context, 90% executed a duplicate refund. INSIDE (6 pages): (1) THE BLIND SPOT — why prompt rules fail in production: a prompt governs what the model can see, not what is true; the full model-by-model failure table. (2) THE COPY-PASTE VALIDATOR — the ~20-line deterministic layer that sits between the model's tool call and execution and checks the action against the source of truth; includes the tool-loop integration and why rejections must go back to the model as tool errors. (3) THE THREE CHECKS — one-time events (functional property), disjoint parties (buyer ≠ staff), closed lists (enum every controlled field) — each with the real-world examples to encode first. (4) THE DECISION RULE — which tools need a validator (moves money / deletes data / messages a human) and which just need a typed schema. (5) THE RECEIPTS — all 392 runs summarized, plus a 10-minute run-it-yourself protocol so you can verify every number. Honest limit included: a validator only guarantees the rules you encode. Independent guide from Hyperautomation Labs.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.