Mult-Agent Strategy in AI: How Anthropics Claude Revolutionizes Search Queries

Anthropic has unveiled technical details about its latest research agent, Claude, which employs a multi-agent approach to enhance and expedite complex search queries.

The system features a leading agent that assesses user requests, formulates a strategy, and subsequently activates several specialized auxiliary agents to conduct parallel information searches. This configuration enables the agent to tackle more intricate requests faster and more thoroughly than a single agent could manage.

During internal evaluations, the multi-agent system outperformed the standalone Claude Opus 4 by 90.2%. The architecture utilizes Claude Opus 4 as the main coordinator and Claude Sonnet 4 as supporting agents.

Anthropic evaluates outcomes using large language models (LLMs) as judges, measuring factual accuracy, source quality, and tool utilization. According to them, this method is more reliable and efficient than traditional assessment techniques. Such an approach allows for the use of LLMs as meta-tools for managing other AI systems.

A crucial performance factor is token consumption: multi-agent deployments utilize approximately 15 times more tokens than standard chats. In internal tests, the amount of tokens used accounted for about 80% of performance variations, with further enhancements arising from the number of tools employed and model selection.

For instance, the shift to Claude Sonnet 4 resulted in a more substantial performance boost than merely doubling the token count in Claude Sonnet 3.7. This indicates that while token usage is important, the choice of model and tool configuration is also critical for performance.

Furthermore, Anthropic claims that in certain scenarios, Claude 4 can identify its own mistakes and refine tool descriptions to improve performance over time, effectively acting as its own consulting engineer.

Anthropic believes that its current multi-agent architecture is particularly well-suited for queries requiring extensive information and allowing parallel processing.

Looking ahead, Anthropic aims to transition to asynchronous task execution, where agents can generate new sub-agents and operate concurrently without waiting for all sub-agents to complete their tasks.

This shift promises greater flexibility and speed but also presents challenges related to coordination, state management, and error handling—issues that Anthropic admits are still being addressed.

On a related note, I’d like to recommend BotHub—a platform where you can test all popular models without limitations. No VPN is required to access the service, and you can use a Russian bank card. Follow this link to claim 100,000 free tokens for your initial tasks and start working right away!

Source