The original title is: "Why unsupervised agent loops self-reinforce errors and how trailing monitors fix them"

Short summary

Unsupervised agent loops (while true) guarantee liveness but not correctness — a wrong turn held in a loop self-reinforces rather than self-corrects. Terminal exit gates fire too late to catch trajectory drift. The fix is a trailing external monitor that runs alongside the loop, scores each iteration, and nudges the agent back on track before it commits to a wrong direction. Apollo Research data shows early drift detection with system-reminder-style redirects is effective but degrades as the agent's confidence grows.

• •while true loops guarantee liveness, not correctness — wrong turns self-reinforce
• •Terminal exit gates fire too late; failures emerge over many turns
• •Trailing external monitors running alongside the loop catch drift early and redirect effectively

Generated with AI, which can make mistakes.