How Should Multiple AI Agents Work Together?

The MAST study, spanning more than 150 tasks across popular multi-agent frameworks, found that roughly 79% of multi-agent failures stem from bad task specification and broken coordination rather than model quality, with inter-agent misalignment alone accounting for about 37% of failures, mostly context lost at handoffs. The study's failure taxonomy indicates that passing summaries and artifacts instead of full histories helps avoid context collapse when agents share context.

Of your four options, handoffs are where most things break, and orchestration matters mainly as the thing that enforces them. The failure taxonomy from the MAST study 150+ tasks across popular multi-agent frameworks found that about 79% of failures come from bad task specification and broken coordination, not model quality. Inter-agent misalignment alone is about 37%, mostly context lost at handoffs. What helps in practice: Shared context matters, but a shared pile of text tends to cause context collapse, so pass summaries and artifacts rather than full histories. Also, this is the same product as in your other recent threads here, so it would help to state upfront that you’re promoting it. People will give you more honest comparisons that way.