A multi-agent pipeline's biggest failure mode isn't any single agent being wrong, it's two correct agents disagreeing about what "done" means

Had a research pipeline where one agent gathered sources and a second agent synthesized them into a summary. Both performed well individually, tested extensively on their own. Chained together, the synthesis agent kept producing summaries missing obvious points the gathering agent had actually found. Turned out the gathering agent considered its job done once it returned anything matching the query, didn't rank or flag relevance strength. The synthesis agent assumed whatever it received was already pre-filtered for relevance and treated every source as equally worth including, which meant weak tangential sources diluted the summary at the same weight as strong direct ones. Neither agent was malfunctioning. They were operating on two different, unstated definitions of what the handoff between them actually meant. This feels like a distinct failure category from either agent just being wrong, which gets most of the attention in agent reliability discussions. It's a contract failure at the interface, the same way two services in a distributed system can both pass their own tests and still break when wired together because neither one's assumptions about the data were ever written down anywhere both sides could check. Started explicitly defining what each handoff is supposed to guarantee, not just what shape the data takes, but what claims the sender is making about it. Fixed more of the output quality issues than any amount of tuning either agent individually.