Multi-Agent Systems: 4 Tests for When One Agent Beats Five

Anthropic researchers demonstrated a system where one lead agent delegated tasks to multiple parallel subagents, outperforming a single agent by 90.2% in internal evaluations. However, this multi-agent approach is more computationally expensive, using about 15 times the tokens of a single agent, which explained 80% of the performance variance on the BrowseComp benchmark.

Consequently, five agents did not win because five heads perform better than one; they won due to the ability to allocate more parallel compute to non-interdependent components, making the design worthwhile only if it pays for itself. The failure analysis of multi-agent LLM systems, published by UC Berkeley researchers, revealed that performance gains are often minimal, with 14 identified failure modes categorized into three groups: system design issues (4 failures out of 10), inter-agent misalignment (~1/3), and task verification (~1/4).

The most significant misalignment issue is a reasoning-action mismatch, where agents say one thing and do another. The authors concluded that reliability requires deeper redesign rather than merely improving underlying models. Before splitting one agent into five, operators should consider these four tests: (1) whether the work actually splits, (2) if the job is worth 15 times the tokens, (3) if each agent can be instructed on what "done" looks like, and (4) who checks the work before it leaves.

These tests emphasize the importance of role design, clear job specifications, individual ownership, and verification, rather than merely scaling up the number of agents. Ultimately, the AI industry is shifting from selecting models to designing roles, with operators needing to think more like managers and less like prompt engineers.