What is inter-agent misalignment?

Inter-agent misalignment is the failure category where agents miscommunicate, lose context during a handoff, or return results that contradict each other, as distinct from one agent simply getting its own task wrong. The MAST taxonomy, built from more than 1,600 annotated traces across seven multi-agent frameworks, names it as one of three root-cause categories alongside specification issues and task verification, and it accounts for 36.94 percent of the failures MAST annotated.

The distinction matters operationally. When a downstream agent’s score drops, the default move is to fix that agent, and that agent is usually fine. The cause sits at the seam: an earlier agent’s prompt changed, or a tool it calls started returning different data, and the output it passed along still reads as valid to anything checking it in isolation. A per-agent grader has no visibility into the handoff itself, so it never sees the seam where the failure actually happened.