Reliability Trap
The Reliability Trap is like a game of 'telephone' played by a computer: even if each person in the line repeats the message perfectly, tiny misunderstandings add up, making the final result completely different from the original, even though every individual step seemed correct at the time.
A failure mode in multi-step AI workflows where the cumulative probability of success drops exponentially with each added link in the chain. Because each individual step maintains a high local confidence score, the system appears functional during unit testing, masking the fact that the aggregate reliability is mathematically destined to degrade.
The agentic analogue of the Eloquence Trap: chained probabilistic steps compound error (0.95^5 = 77%; one in four outcomes wrong at five steps), and because each step examined alone appears sound, the failure is invisible until the end. A mathematical certainty to design for, not a bug to fix. (C4AIL.)