A developer implemented a multi-agent reasoning system using Anthropic's Claude extended thinking API to improve accuracy on complex math problems by having two independent agents reason separately and then resolving disagreements with a third agent. This approach, called self-relay, achieves a 2.5 point accuracy gain over single-agent reasoning at about 2.7x token cost, significantly more cost-efficient than a prior relay-structured approach. The system uses confidence scores to gate when to escalate to the more expensive multi-agent reasoning, reducing average cost overhead to about 1.3-1.5x single-agent. The method is applicable to domains where the cost of wrong answers is high, such as legal document review, code security audits, medical triage, and financial analysis.

Jun 3, 2026, 7:35 PM