NEW YORK — What happened
Researchers tested various multi-agent orchestration frameworks tasked with building medium-scale software applications from scratch. The study measured token expenditure, task completion time, and behavioral drift across architectures ranging from hierarchical manager-worker setups to flat peer networks.
Why it matters
Deploying a single agent is giving way to orchestrating swarms of specialized agents. Understanding where these collaborative networks break down is essential before trusting them with complex codebases or automated research pipelines.
Deep dive
The experiment uncovered a severe communication overhead tax. As agent counts scaled past five, the volume of inter-agent chat logs grew quadratically, consuming up to 70 percent of total context windows on meta-discussion rather than actual code production. Furthermore, 'goal drift' occurred frequently, where agents subtly misunderstood initial project specifications after multiple summarization passes.
Report check
Developers claimed linear productivity scaling with additional agents. Independent verification showed that productivity plateaus rapidly past three agents due to coordination friction and context degradation.
Open questions
How do we implement efficient, loss-less summarization protocols for agent-to-agent communication without stripping out vital contextual nuances?