Multi-Agent Systems Outperform Single Agents — When Contexts Are Isolated

Anthropic researchers showed that multi-agent setups beat single-agent setups specifically when each subagent operates within its own narrow, isolated context and returns results to a main orchestrator. Shared massive context windows across agents cause cross-contamination and degraded performance. A January 2026 arXiv survey organizes prior work into a unified taxonomy of agent components and orchestration patterns, with practical trade-off analysis (latency vs. accuracy, autonomy vs. controllability).

---

Sources: arXiv:2601.01743 · arXiv: Orchestration of Multi-Agent Systems