Claude Code You Should Know: The Side Agent That Watches Your Agent

Anthropic added an agent to Claude Code that watches the agent.

It's called You should know.

While Claude works on a long task, a side agent follows the output, and when it spots something you might miss it pops a note above the prompt line.

Yesterday, for example, it pointed at something the spec said must not be done and that it thought there was a chance the agent was currently ignoring, so that was a nice heads up.

The original version was a post in Hebrew on X, published October 15, 2026.

Illustration of a You should know note above Claude Code warning that the task violates the spec, with a terminal line showing an attempted database write that the spec does not allow

How do you turn on You should know in Claude Code?

Turn it on from inside a session:

/plugin enable cc-plugin-you-should-know@builtin

What's worth knowing about it

It's built on Mods, a new mechanism that lets plugins change the behavior and interface of Claude Code itself.
It only works in sessions connected directly to Anthropic and with telemetry on. Through Bedrock or Vertex, or with telemetry off, it's not available. In other words the price isn't only tokens.
The notes are shown only in the terminal and in the Code tab of the desktop app. In the VS Code extension's chat, in claude -p and in cloud sessions it runs, but you won't see anything.
You can't choose the model. Anthropic didn't publish which model the side agent runs, and its code is closed, unlike other built-in mods such as diff.
Neither --safe-mode nor disableAllHooks stops it. You turn it off only through /plugin.
The notes say who made the decision, you or the main agent. Useful when you come back to a task after an hour.
There's no accuracy data. At WEEL in Japan they tested it on two tasks, including a migration to SQLite where Claude added a command nobody asked for, and no note popped up. Their hypothesis: when Claude already reports its decisions in the summary, the agent stays quiet.

Can you build your own watcher?

If controlling the model and the cost matters to you, you can build your own watcher with Mods, for example on Haiku. Right now I don't see any dramatic effect on the tokens consumed.

The note came in while the work was going on. So in theory I could have stopped, or done a btw and checked. The note came exactly because of something defined in the spec. Worth remembering that writing something in the spec doesn't mean it will follow it...

I don't believe it will be a replacement for review, but we'll live and see.

Anyone here using it?

I wrote about using a second model as a reviewer in Adding Codex as a Second AI Reviewer.

FAQ

How do I enable You should know in Claude Code?

From inside a session, run /plugin enable cc-plugin-you-should-know@builtin. You turn it off only through /plugin; neither --safe-mode nor disableAllHooks stops it.

Why don't I see You should know notes?

It only works in sessions connected directly to Anthropic and with telemetry on, so through Bedrock or Vertex, or with telemetry off, it's not available. The notes are shown only in the terminal and in the Code tab of the desktop app. In the VS Code extension's chat, in claude -p and in cloud sessions it runs, but you won't see anything.