Towards self-driving codebases · Cursor
We’re making a part of our multi-agent research harness available to try today in preview.
AI Architecture Blog Post
Summary
Simon Willison's analysis of Anthropic's multi-agent research system architecture, where a lead agent spawns parallel subagents for breadth-first research tasks, outperforming single-agent approaches by 90.2% at the cost of 15x token usage.
Highlights
- Multi-agent system with Claude Opus 4 lead and Sonnet 4 subagents outperformed single-agent Opus 4 by 90.2% on internal research eval
- Multi-agent systems use approximately 15x more tokens than standard chat interactions, a significant cost trade-off
- Architecture excels for breadth-first queries requiring parallel exploration of independent research directions
- Agent defined as LLM autonomously using tools in a loop; subagents compress findings from parallel context windows
- Simon Willison notes this cured his skepticism about multi-agent LLM systems after seeing the detailed engineering writeup
auto-generated
Context
Audience
AI engineers, LLM application developers, and technical decision-makers evaluating multi-agent architectures
DomainLarge Language Model Architecture
Formatblog post with technical analysis
Accessfree online
Source Overview
OK, I’m sold on multi-agent LLM systems now. I’ve been pretty skeptical of these until recently: why make your life more complicated by running multiple different prompts in parallel when …
Topics
Related
Claude ResearchClaude CodeAnthropic engineering blogLLM agent orchestrationToken cost optimization
