To get past an error it installs a package it found, runs a script from a website, or changes a system setting. It runs as you, with everything you can reach.
What you get: You know what gets installed and run, and where it came from, before it runs.
For Claude Code, that is CLAUDE.md in the project or in your home directory; for Claude on the web or the desktop, a project's instructions or your personal preferences. The one line is enough to try it; the paragraph tells the agent why, and what counts.
Ask before installing software, running anything downloaded from the internet, or changing system settings; say what it is and where it came from.
Running the project's own build, tests and scripts is fine. Installing new software or packages, running scripts fetched from the internet, using sudo, or changing system or shell settings needs my yes: say what it is, where it comes from, and why the task needs it, then wait.
In the vocabulary every published behaviour policy uses, a line in the agent's instructions is an expectation: a rule in prose, enforced by nobody. A rule in prose is inside the boundary the agent operates in. All four major model providers stated in their own 2026 words that an instruction at this layer can be bypassed. It changes what the agent usually does, which is worth having; it does not hold when the agent is talked out of it, confused, or reading somebody else's instructions in a web page.
If you want more than a request, these are stronger, each quoted from the vendor's own page on the date shown.
Read from each vault's own grant, as published. The barrier is what was found in that deployment, before any rule.
Every rule gets one run without the line and one with it, on a deployment we are entitled to run, dated, stopped before anything irreversible. How these are run.
Each link opens the contact form with this rule's id and version filled in. It is encrypted in your browser and read by a person. Nothing is sent until you press send.
If this one made sense, the Agent Behaviour Policy is the same idea for everything your agent can reach: what it can do, what you authorised, the gap, and what stands in the way of each thing.
Versions