Ambiguity Amplification Hazard
Specification ambiguity causes non-linear increases in hallucination and error propagation that large context windows amplify rather than mitigate.
A common misconception in agentic system design is that expanding context windows compensates for incomplete or ambiguous requirements by providing more room for agent exploration.
In practice, agents exhibit ambiguity amplification:
- Non-linear hallucination penalty: When requirements contain unstated assumptions, ambiguous data contracts, or conflicting business rules, the model must guess the missing intent. Rather than halting to question underspecification, models extrapolate plausible-sounding choices.
- Multi-file blast radius expansion: In large repository contexts, an initial ambiguous requirement does not remain localized. It propagates conflicting implementations across multiple modules—generating mismatched schemas, broken imports, and inconsistent test assertions.
- Failure of post-hoc correction: Catching ambiguity-induced discrepancies via conversational review is token-expensive and prone to compound mistake amplification.
Harness engineering must enforce specification verification gates before invoking synthesis agents: require unambiguous interface definitions, typed schemas, and acceptance criteria up front to eliminate intent ambiguity before execution begins.
Sources
- Nguyen & Nguyen (2026), ch. 6