Member-only story
Chapter 1 covered the agentic loop. Chapter 2 covered coordinators and subagents. Chapter 3 covered tools and MCP. Chapter 4 covered structured output. Chapter 5 covered CI/CD and unattended safety.
This is the last chapter. It closes a gap from Chapter 1: every tool call in the loop can work perfectly — the right file gets read, the right API gets called, the right answer comes back — and the agent can still fail. Not because anything broke. Because something it needed to remember two turns ago quietly isn’t there anymore.
This maps to Domain 5, Context Management & Reliability — the last domain on the exam, worth 15%. It’s built around the Customer Support Resolution Agent scenario: an agent with four tools — get_customer
, lookup_order
, process_refund
, escalate_to_human
— that's supposed to resolve at least 80% of conversations on its own. This chapter isn't about the model getting dumber. It's about what happens to information after the model produces it. That's where most of the failures actually live.
Why a working loop can still lose the plot
Every tool call adds more text to the conversation. Here’s the part people forget: the Claude API has no memory of its own. There’s no session sitting on a server somewhere remembering what happened three turns ago. Every single request has to carry the entire conversation again, from the start. So there’s no such thing as quietly forgetting one old…