Daily research
Autonomous database agents

Why Autonomous DDL Needs Rollback Contracts

A safe autonomous Postgres agent should prove how it backs out before it asks to change a production database.

/Hypothesis/Database Agents

Database owners evaluating autonomous operations

Research question

What makes an autonomous database change reviewable by a human operator? The hypothesis is that rollback classification is part of the minimum viable proof.

Method

For each action class, record lock behavior, dry-run evidence, rollback SQL when available, and the checks QueryRook would run after the rollback.

Operator use

The review screen should answer the uncomfortable question first: if QueryRook is wrong, what exactly happens next?