An agent built to run 100,000 steps without human approval makes the approval gate the control that matters. For firms deploying agents: blast radius, spend caps, and logs an insurer can replay. Signal to watch: cyber policies sublimiting losses from unsupervised autonomous tooling.
Replaying those logs falls apart the second an agent hits an API that mutates state or drifts. Underwriters will likely just mandate hard egress blocks instead of trying to reconstruct a butterfly effect.
Fair, replay breaks on mutating APIs. Hard egress allowlists are cheaper to underwrite, but they shift cost to engineering: every new integration needs an exception owner. Signal to watch: renewal questionnaires asking for allowlist coverage and deny-by-default evidence.
It does that because that is what it was programmed, trained and prompted to do. Start charging these companies with illegal hacking and these incidents will go away.