References: Many agents working together: multi-agent systems

Source material

Source curriculum (structural mirror, cited as further study):
• Microsoft, "Multi-agent design patterns" (AI Agents for Beginners, Lesson 08)
  Author: Microsoft Cloud Advocates
  Lesson page: https://github.com/microsoft/ai-agents-for-beginners/tree/main/08-multi-agent
  License: MIT
• Berkeley CS294/194-196, "Compound AI Systems & the DSPy Framework" (Lecture 5)
  Speaker: Omar Khattab (Databricks)
  Course page: https://rdi.berkeley.edu/llm-agents/f24
  Video: https://www.youtube.com/live/JEMYuzrKLUw
  License: YouTube standard (link-out only)
Clawdemy's lessons are original prose that follows the pedagogical arc of these
sources. We do not reproduce or transcribe them; we cite them as recommended
companions. All rights to the original materials remain with their authors.

Note: Microsoft Lesson 08 presents multi-agent advantages without the
coordination-cost downsides. This lesson supplies that counterweight
deliberately, for an honest fit-not-ranking treatment.

Read or watch this next

• Multi-agent design patterns (Microsoft) by Microsoft Cloud Advocates. The practitioner version, with coordination patterns, agent communication, and a worked refund-process scenario, plus runnable samples. MIT-licensed. Strong on the advantages and the building blocks; pair it with this lesson’s coordination-cost section for the full tradeoff.
• Compound AI Systems & the DSPy Framework (Omar Khattab, Berkeley CS294 Lecture 5). The broader framing: modern AI applications as systems of composed components, not a single model. The lens that makes multi-agent one case of a general design idea.

Going deeper on compound systems

A short, durable list.

• The Shift from Models to Compound AI Systems (Zaharia et al., Berkeley AI Research, 2024). The blog post that named and argued for the compound-systems framing. The clearest single statement of why systems-of-components, including multi-agent, are where capability is increasingly built.
• Model Context Protocol. An open standard for connecting agents to tools and data, useful infrastructure when several agents need to share the same tools and sources.

Adjacent topics

Where this leads inside this track.

• Agents that self-check: metacognition. The next lesson. A way to make a single agent more reliable without adding a second one: a reflection step where the agent checks its own work before committing.
• Planning: breaking a goal into steps. The previous lesson. A plan’s sub-tasks are what a multi-agent system divides among specialists, so planning and multi-agent design are two halves of the same decomposition question.
• The tool-use design pattern in depth. Earlier in the track. The reason specialization helps: an agent with a small, well-described toolbox picks more reliably than one with too many tools, which is the system-level case for splitting into specialists.