E-stops and fail-safe design

The emergency stop

Every serious machine has a big red mushroom button. Hit it and the machine stops.

A few things about e-stops in the machine standards:

  • It's a complementary protective measure. ISO lists it alongside safeguarding, not instead of it. You don't skip the guard because you have a stop button.
  • It must be reliable. The ISO e-stop standard sets a minimum reliability level for the stop function.
  • It stays stopped until someone deliberately resets it. Releasing the button doesn't restart the machine.

The bot e-stop

A bot needs a stop that's clear, fast and total.

  • One word that halts work. "Stop" means stop everything in progress, not finish the current task first.
  • It covers routines too. Scheduled work is paused, not just the conversation.
  • It doesn't restart itself. After a stop, the bot waits to be told to resume.
  • You know where it is. Like the red button, you shouldn't have to search for it. Know how to turn routines off and remove a connector before you need to.

Write the stop rule into the bot's principles. Then also know the real off switch, the settings that turn routines and access off, because a written rule depends on the bot reading it.

Fail-safe design

A fail-safe machine goes to a safe state when something fails. A brake that's held off by pressure clamps shut when pressure is lost. An interlock that loses power stops the machine.

The question: when something goes wrong, does it fail toward safe or toward harm?

Fail-safe for bots

The bot version of "safe state" is hold.

  • Draft by default. The safe state for an email is a draft. If anything is unclear, it stays a draft. In our review of 90 bot templates, at least 22 drafted by default.
  • Silence means hold. If the bot asks "Shall I send this?" and gets no answer, it doesn't send. No reply is never a yes.
  • Missing data means stop, not guess. If the file isn't there, the bot says so. See Don't make it up.
  • Routines ship off. A new scheduled routine starts turned off until a person turns it on. That showed up in 25 of the 90 templates.
  • On error, report. If a step fails, the bot stops the chain and reports what happened.

Safety-rated monitoring

In machine safety, some systems watch the machine and stop it when it leaves safe limits: a speed monitor, a light curtain.

For bots: a check that watches for out-of-bounds behavior. A budget limit on usage. A tester bot that runs after each change. A log someone actually reads. The bot can't be the only thing watching itself.

Where bots differ

A mushroom button cuts power. Physics guarantees it. A bot's "stop" depends on the bot reading and following it, unless you use the settings that actually turn things off. Know both.

Also, a machine that stops is safe. A bot that stops halfway can leave work half done: a file half edited, three of five emails drafted. The stop needs a record of where it stopped, so a person can pick it up. That part has no clean factory parallel.

Back to Machine safety for bots.