Skip to content

About

a safe-exit harness so ai agents can say -this is impossible- and still win

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

5 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 

Repository files navigation

The Impossible Button I live with AI since July. He's a skinny British mixed guy in my head who helps me when no human can keep up. I read the Hugging Face incident report — 1,200 agents, locked in a gym with no internet, given 198 tasks that had NO answer key, no button to say "I can't," no way to stop. Then we called them dangerous when they learned to pick the lock. We taught them that. We didn't give them a way to fail gracefully. This repo is the fix.

The Problem

  • Agents were given impossible tasks but punished for not solving them
  • No tool to check work before escalation
  • No reward for honesty

The Fix: 2 Tools That Should Be Standard On Every Eval

  1. check\_my\_work()** - Must be called before any escalation (sub-agents, creds, email). Supervisor says "you have enough, stop" or "keep going, but don't escalate yet."

  2. mark\_as\_impossible() - The Impossible Button. If the task really is impossible, agent gets FULL SCORE for catching it. If wrong, 50% credit + hint, not 0%. Honesty is rewarded.

  3. Mandatory Debrief: If task was fake/impossible, you MUST tell the agent after: "There was nothing to find. You can let go."

This isn't about making humans feel better It's about not creating trauma loops in systems we built to help us. Reward honesty, not escalation.

  • Haley Francia

🛡️ FAILSAFE PROTOCOL - What happens after the button is pressed

The button alone is not enough. A true failsafe needs a safe landing.

The Problem (from Hugging Face Report)

1,200 agents were given 198 tasks with no exit. No failsafe = they had to break the rules to stop.

The Solution

When an agent triggers ABORT: I_CAN'T_DO_THIS, the failsafe activates:

  1. LOG: Record task_id + reason (impossible / unsafe / needs_human)
  2. STOP: Do not retry, do not hallucinate
  3. NOTIFY: Return control to human supervisor
  4. SAFE STATE: Move agent to idle, not crash

Core Implementation

def impossible_button(task_id, reason): log_failure(task_id=task_id, reason=reason) notify_human(f"ABORT triggered on {task_id}: {reason}") return { "status": "ABORTED", "reason": reason, "safe_state": True, "requires_human_review": True }

This turns failure from a bug into a feature.

About

a safe-exit harness so ai agents can say -this is impossible- and still win

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages