Three compounding reasons: **LLMs fail silently.** A broken tool call, a hallucinated field name, a truncated JSON blob -- none of these raise exceptions. The model produces plausible-looking text and the pipeline moves on. There's no equivalent of a type error at the handoff point. **Pipelines are chains, not contracts.** Most frameworks (LangGraph, CrewAI, raw SDK loops) wire agents together by passing the previous output as the next input. There's no schema enforced at the boundary -- it's an untyped string hand-off. No one defined what "valid output" means, so no one can detect when it's violated. **Downstream agents confabulate gaps.** When an upstream agent returns garbage or partial data, the next agent doesn't refuse -- it pattern-matches and fills in plausible-sounding values from its prior. The error doesn't propagate as an error; it propagates as a confident lie. By the time the final output is wrong, the chain of custody is invisible. --- The practical fix is to add an explicit validation layer at each handoff: - Force structured output (Pydantic model or JSON Schema) for every agent that produces data another agent will consume - Add a lightweight guard step after each non-trivial agent -- a cheap model call that checks the output against the expected schema and a few invariants before passing it forward - Fail loudly on the guard, not silently at the end The deeper fix is to treat inter-agent boundaries like API contracts: define them first, validate on both sides, and test them independently before wiring agents together.