Why is off by default the right setting for an agent that watches another agent?

Because the people with the most data on how often a side agent has something worth saying chose it. Of the six mods Anthropic ships inside Claude Code, You Should Know is the only one that spends the user's attention and the only one off until the user turns it on, even though McAteer reports it saved him from discarding uncommitted work in about 45 files. The reader opts in.

What is a budget line for an agent interface?

A share of the person's attention, set before the trigger condition is written: how many times per session, or per day, the surface may interrupt. A trigger such as 'after any turn with eight or more tool calls' is not a budget, and in McAteer's case it matched 99 of 111 turns. Jakob Nielsen's point is that the interruption budget belongs to the user and every agent draws on the same account.

What should be counted from day one?

Two columns: turns where the person acted on what the surface said, and turns where it spoke. McAteer published the second column. The first is the one that says whether the surface earned its place. In my fleet the equivalent was counting what read each agent's output; the orphans had zero references from outside the fleet, the agents that mattered had three or more.

When Should an AI Agent Interrupt You?

When should an AI agent interrupt you?

THE SHORT ANSWER

Only inside a budget the person set, and only after the surface has earned it. Dan McAteer's October 21, 2026 Latent Space post on Claude Mods supplies the number: his Next Steps Supervisor fired on 99 of his 111 turns since October 15, and he started skimming past it, while Anthropic ships its own watcher mod, You Should Know, off by default. My April audit of 39 appointed agents found 13 firing in one morning window with nobody planning it and 13 orphans nothing read. So three lines before an agent-facing surface speaks: off by default, a budget line for how many times per session it may interrupt, and a count of turns where the person acted on what it said against turns where it spoke. The tie-breaker is the disclosure test from the design handbook: it earns its place only if it changes what a reasonable reader does next.

Who writes the interface is the wrong axis. The default and the budget are the design, and the first programmable attention interface shipped with the same failure as the appointed fleet: every piece speaking because it can.

The three lines, McAteer's own test, and the fleet numbers are in The Mod Fired on 99 of 111 Turns. Off by Default Is the Design.. The April audit (13 of 39 in one window, 13 of 39 orphans, 15 of 44 today) and the charter version of the same three lines are in 44 Appointed Agents, One Attention Budget: Three Lines on the Charter. The disclosure test comes from Disclosure Without Overwhelm in the design handbook.

Last reviewed 2026-10-21 · 1 min read

When should an AI agent interrupt you?

SOURCES

THE LONG VERSION

RELATED ANSWERS