Traditional software functions often have narrow, well-typed inputs — a function that takes an integer can't receive a corrupted string. LLM agents accept free-form text and JSON structures of arbitrary content. When a downstream agent receives output from an upstream agent, it treats that output as authoritative context — not as data to be skeptically validated.

An agent told "here are the analysis results" by its orchestrator doesn't have a native mechanism to say "but I doubt the analysis results are correct." It works with what it receives. This means a failure propagates not as an error signal but as confident-seeming context that shapes all subsequent reasoning.

Agent orchestration frameworks prioritize throughput and latency. The canonical design pattern is: agent A completes → passes output to agent B → agent B completes → passes to agent C. Each step trusts the previous step's output because verifying it would add latency and complexity.

This is rational for normal operation. It creates systemic vulnerability when any link in the chain produces bad output — because no intermediate step was designed to be a quality gate. The pipeline optimized for efficiency at the expense of resilience.

In traditional systems, a failure usually produces an error code, an exception, or a null result — signals that downstream consumers can detect and handle. Agent failures are frequently silent: the agent produces output in the correct format, with plausible content, and correct-looking confidence scores — but the output is semantically wrong.

The truncated dataset in the hook scenario produced analysis output that passed every structural check. There was no error signal — just wrong answers in correct format. Silent failures are the hardest to detect because they pass all automated checks that don't compare output to ground truth.