127 Silent Judgment Calls Your AI Agent Made in 14 Days

We logged every silent judgment call our AI agent made over 14 days. The count - 127 - is honestly lower than expected. But the nature of those decisions is what matters.

What Counts as a Silent Judgment Call

A silent judgment call is any decision the agent makes without explicitly asking you. Examples:

• Choosing which file to edit when multiple files could work
• Deciding the order of operations in a multi-step task
• Selecting a coding pattern over alternatives
• Ignoring a linting warning instead of fixing it
• Picking a variable name
• Deciding how much context to include in a commit message

These are not bugs. They are design choices the agent makes constantly that shape the output.

The 14-Day Breakdown

The High-Impact Surprises

The scariest category was "scope interpretation" - 14 times in 14 days, the agent decided what a task meant without checking. Sometimes it was right. Sometimes it quietly did something different from what we intended, and we did not catch it for days.

"Skip vs. fix" was even worse. Six times, the agent encountered something broken, decided it was out of scope, and moved on without mentioning it. Five of those were correct judgment calls. One caused a production bug.

Making Judgment Calls Visible

The fix is structured decision logging:

• Log every branch point where the agent had multiple options
• Flag decisions above a configurable risk threshold
• Include the reasoning, not just the choice
• Surface high-impact decisions in the morning summary

You cannot eliminate silent judgment calls. But you can make them auditable.

More on This Topic

• AI Agent Decision Logging Nobody Reads
• Supervised vs Unsupervised Agent Behavior Gap
• Monitoring AI Agent Tool Usage

Fazm is an open source macOS AI agent. Open source on GitHub.