COMMUNICATION AND PROTOCOLS BETWEEN AGENTS

CHECK YOUR UNDERSTANDING • What is the main benefit of a message bus over direct agent-to-agent calls? ◦ Answer: Loose coupling, since publishers and listeners need not know about each other, plus a natural place to log everything. • How do MCP and A2A differ? ◦ Answer: MCP mainly connects an agent to tools and data, while A2A mainly connects an agent to other agents. • Describe a deadlock in agents. ◦ Answer: Two agents each wait for the other's result, so neither ever proceeds. • List two defenses against deadlock or runaway loops. ◦ Answer: Timeouts with escalation to a human, and limits on handoffs, rounds, time, and budget, or cycle detection. • What does a meta-harness provide? ◦ Answer: Shared registry, budgets, permissions, logging, policy enforcement, and kill switches across many agents. PART ELEVEN: MEMORY, KNOWLEDGE, AND RETRIEVAL An LLM, on its own, remembers nothing between runs. Every conversation starts with a blank slate, and everything the model knows about your situation must be placed in front of it again. Part Eleven is about how a harness gives an agent memory, how it fetches the right knowledge at the right moment, and how it keeps that knowledge organized and fresh. Chapter 46 covers the types of memory. Chapter 47 covers Retrieval-Augmented Generation, the technique for grounding answers in real documents. Chapter 48 covers knowledge management, including skills, notes, and keeping information from going stale.