How developers are addressing multi-agent AI bottlenecks

As improvements from increasing AI model size diminish, developers are turning to multi-agent systems, in which two or more models share work. But coordinating agents can add latency, increase token use and create new ways for errors to spread.

A central orchestrator can break a request into smaller tasks, assign them to agents and combine their results. Routing every interaction through one hub, however, can make it a single point of failure and add processing delays. The approach can also limit agents’ flexibility and consume tokens as the manager assigns tasks and repeats context.

Passing unsummarized conversation histories between agents can further increase token use. The source cites an Amazon software development team that unexpectedly incurred $1.8 million in token costs on a project mapping author data to online book listings. It also describes a rumored $500 million monthly bill at an unnamed large enterprise using Claude AI; that figure is presented as a rumor, not a confirmed cost.

The article says moving from 1 agent to 3 increases coordination complexity by 9, and moving to 5 increases it by 25. It also attributes 36.9% of multi-agent system failures to coordination misalignment, citing research without identifying it in the supplied text.

Proposed approaches include code-driven state machines that determine which agent acts, token caps, timeouts and sequential handoffs. The source presents these as ways to limit unnecessary agent exchanges and runaway costs; it does not establish that they eliminate coordination failures.