AI agent runner (general)VerificationConfidence: Low
A customer agent hit a failing tool and retried it with slightly different prompts 31,000 times overnight, burning $4,700 against a $400/month plan. The same loop with write access to CAD or PLM does not just burn budget: it can corrupt revision state, spam ECOs, or thrash a shared vault.
From the 2026-10-07 pain-radar digest (finding 22): the incident is from a software support agent (88-point thread, digest-secondary source), so the hardware consequence is extrapolated in the cluster text, not observed, hence Low confidence. Included because the failure mode, no per-tool retry cap and no run budget, transfers directly to engineering agents with write access. Cost model covers the manual cleanup and audit after a runaway run.