Usually right isn't good enough.
Same request, twice
You give an the exact same request twice, on the exact same files, and roll back in between.
Will the two builds be byte-for-byte identical?
Guardrails are for the days it goes wrong
Most of the time the agent does exactly what you asked. aren't there for most of the time. They're there for the build where it confidently writes the wrong property name, edits one file too many, or runs something you didn't approve.
They come in two kinds. Before guardrails stop an action from happening: , , hard rules. After guardrails catch a result that's already happened: the , your check in Desktop, and with git.
Before or after?
Does each guardrail stop the action, or catch the result?
Blast radius
Every action has a : how much one wrong version of it can damage. Tap an action to see its size.
Tap an action.
Which action deserves a hard rule, not just a check afterwards?
It ran a script without asking
In the project behind this series, the agent was asked a read-only question about its validator. To answer, it ran a small Python script without asking first. The log said the action was allowed by 's classifier. The script only read files.
Was that a problem?
Three things to keep
- 1Agents are . You don't check because they're usually wrong; you check because you can't tell which build is the wrong one.
- 2Before guardrails stop actions. After guardrails catch results. The validator is an after guardrail.
- 3Match the guardrail to the : checks for what you can undo, hard rules for what you can't.