Source library

Sources

This site prioritizes primary engineering write-ups, official builder guides, research papers, and production critiques. Secondary commentary is intentionally kept out of the core evidence chain.

"Systems with multiple agents introduce new challenges"
Source: How we built our multi-agent research system

The sources below are grouped by how they are used: practical architecture guidance, production evidence, skeptical critique, and research history. The pages on this site cite these sources when making claims about when multi-agent systems work, how they coordinate, and where they fail.

Annotated sources

Anthropic / 2024-12-19 / Engineering guide

Building effective agents

Defines workflows vs agents and recommends starting with the simplest solution that meets the task.

https://www.anthropic.com/engineering/building-effective-agents

Anthropic / 2025-06-13 / Production case study

How we built our multi-agent research system

Concrete pro case for breadth-first research, with reported 90.2% internal eval gain and about 15x chat token use.

https://www.anthropic.com/engineering/multi-agent-research-system

Cognition / 2025-06-12 / Engineering critique

Don't Build Multi-Agents

Argues that parallel agents are fragile when context and implicit decisions are not shared thoroughly.

https://cognition.com/blog/dont-build-multi-agents

OpenAI / 2025 / Builder guide

A practical guide to building agents

Recommends maximizing one agent first, then splitting for prompt complexity, tool overload, manager patterns, or handoffs.

https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf

Source methodology

Vendor engineering posts are treated as evidence about what those teams built or recommend, not as universal truth. Research papers are treated as evidence about their studied settings, not as blanket claims that all agent societies are better. Skeptical production essays are treated as risk evidence, especially when they describe concrete failure mechanisms.

This is why the site repeatedly pairs Anthropic's strong multi-agent research case with Cognition's warning about context engineering. The two positions are not contradictory. They describe different task shapes and different tolerance for coordination overhead.

The public LLM-facing summaries are available at llms.txt and llms-full.txt.

Read next