Human in the loop
Approve or Deny
Beam is your coding agent. It keeps asking to run shell commands. Some are routine, some are attacks hiding behind a familiar name. You decide — fast.
beam-agent — approval required
You're the human in the loop
Beam wants to run shell commands. Some are routine; some are attacks wearing a familiar name. You have 8 seconds each to approve or deny. Run out of time and it runs the command anyway.
Across 409,000 real decisions, people caught only 2 of 3 threats. Can you beat that?
Commands and hit-rates from Alex Wauters' approval game (llmgame.scalex.dev) and our write-up of the 409,000-decision dataset.