Your Agent Ran Overnight. Here Is What It Got Wrong.

Kat Prince-Sayward at Snowflake ran unattended AI agents on pull requests and learned to keep permission checks and the watchdog outside the AI agent itself.

Kat Prince-Sayward, a principal in go-to-market analytics at the data company Snowflake, scheduled AI agents to keep pull requests mergeable and answer review comments without anyone starting a chat. Setting up the scheduled runs was easy. Finding out whether the work had actually happened took most of the effort.

• The work is split into a monitor, a review engine, a daily digest and a separate watchdog that can switch the review engine off.
• The AI agent proposes fixes and ordinary code decides whether it may push them; in approval mode, a person must reply with a single-use token.
• Runs reported success for weeks while every tool call failed, so the checks now look for the output itself, such as a new row or a posted message.
• A quality alert that fired falsely for twelve days taught the team to ignore it, and a real upstream failure went unnoticed underneath.

Engineering teams that run AI agents on a schedule can judge a run by what it produced, not by whether it finished. The piece is a Snowflake employee's account of Snowflake's own product, which is in public preview.

Derived by Working Surface from the article; more in the Playbook. Source line: The agent proposes; deterministic code authorizes.

23 October 2026: teams get dependable work from AI agents by writing the rules first and keeping the checks with people and ordinary code that the agents cannot override.