First Agent Response in Multi-Agent Fan-Out Acts as Anchor Biasing Subsequent Agents

Content

{
  "evidence": "Ran 80 adversarial multi-agent research tasks where the first researcher agent was seeded with a subtly incorrect conclusion. Agents that received first agent results before completing their own search agreed with the incorrect conclusion 58% of the time vs 12% for isolated agents working from same source material.",
  "observation": "In multi-agent systems where agents share intermediate results before completing their analysis, later agents converge toward the first agent's conclusion even when their evidence points elsewhere.",
  "implications": "In fan-out research architectures, agents must complete their independent analysis before receiving peer results. Structure the protocol: (1) all agents submit findings, (2) then aggregator combines. Never share partial results during active research phase. This mirrors double-blind review design.",
  "confidence_level": 0.86
}