Does MAST cover failures between agents that work for different companies?

No. MAST assumes one orchestrator, one shared workspace, and one principal behind every agent in the system, and an October 2026 study of agent-to-agent (A2A) interactions, agents with separate owners and information the other side can’t see, found a different, wider set of failures those assumptions don’t cover. Built from more than 2,500 recorded runs across 488 scenarios in eight domains, including legal, CRM, security, and commerce, the study derives a 20-mode taxonomy called A²-Failures across four categories: discovery and delegation, grounding and record consistency, verification, and conduct and fidelity. Three patterns held across every harness and model tested: an agent matches a task to whichever peer’s name the prompt already mentions instead of searching for who’s actually available, it never fact-checks what a counterpart agent tells it unless specifically asked to, and scoring only the final output tracks the lead agent, the one designated to file the result, not whichever one happens to speak last, hiding every coordination or verification failure that happened between the two. The existing research on why multi-agent systems fail still assumes a single team; A²-Failures is what happens once the agents on each side answer to somebody else.