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"
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
Microsoft Research / 2024-08 / Research publication
AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation
Documents a framework for composing conversable agents with LLMs, tools, code, and human input.
https://www.microsoft.com/en-us/research/publication/autogen-enabling-next-gen-llm-applications-via-multi-agent-conversation-framework/
Du et al. / 2023-05-23 / arXiv paper
Improving Factuality and Reasoning in Language Models through Multiagent Debate
Studies multiple model instances that propose, critique, and converge over debate rounds.
https://arxiv.org/abs/2305.14325
Hong et al. / 2023-08-01 / arXiv paper
MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework
Frames cascading hallucinations as a core risk and uses SOP-style workflows to reduce errors.
https://arxiv.org/abs/2308.00352
Li et al. / 2023-03-31 / arXiv paper
CAMEL: Communicative Agents for Mind Exploration of Large Language Model Society
Introduces role-playing communicative agents for studying cooperative behavior in multi-agent settings.
https://arxiv.org/abs/2303.17760
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.