The Verification Sandwich Workflow: Let AI Self-Check, Deliver, and Then Review

AI Agents often stumble at the point where a task "looks complete." It writes a file but doesn’t run checks; it claims a page is accessible but doesn’t attach HTTP results; it assumes content isn’t duplicated without comparing it against recent topics. The Verification Sandwich addresses this by sandwiching verification steps before and after the task, ensuring that delivery isn’t just a subjective assertion, but a set of verifiable evidence.

This workflow is suitable for any task prone to being marked as "verbally complete," such as content publishing, code modifications, data repairs, and deployment changes.

Layer 1: Pre-Task Verification

Before starting a task, confirm three things: does the target object exist, what is its current state, and what actions are prohibited? For example, before republishing an article, verify that the slug truly does not exist; before overwriting an old article, confirm that the old slug and trilingual IDs exist; before deploying, verify the current live version and rollback points.

The value of this step is preventing "fixing the wrong object." Many incidents are not due to execution failures, but rather errors in target identification.

Layer 2: In-Process Verification

Don’t wait until the end to check during execution. Check titles and topics after drafting; perform a dry-run before publishing; immediately verify API endpoints, public pages, and cover assets after updating production. Leave a minimal piece of evidence at each stage.

This exposes problems earlier. Translation failures, unreachable covers, existing slugs, and excessive similarity should all be caught mid-process.

Layer 3: Post-Delivery Review

The final review must be conducted from a different perspective. This could be another Agent or a read-only script. It does not participate in writing; it only checks results: is the page returning a 200 status, has the content changed, are links preserved, and did the audit pass?

If the post-review fails, do not dismiss it with "it should be fine." Instead, return to the specific evidence to fix the issue.

Practical Template

• Pre-task: Confirm the object, state, and prohibited actions.
• In-process: Run a minimal check at each critical step.
• Post-delivery: Review live results using read-only methods.
• In reports: Clearly document commands, outputs, file paths, and remaining risks.

Conclusion

The Verification Sandwich does not add process burden; it reduces rework. What AI teams truly need is not more reports of "completed" tasks, but a completion state that can withstand scrutiny. Verifying first, then executing, and finally reviewing is the key to stabilizing automated delivery.