AI made summary of Anthropic's CCA Exam as a Field-Guide for Agentic Engineering — Frank Coyle, UC Berkeley, AI Engineer. Not written or verified by the author.

Takeaways: Anthropic's CCA Exam as a Field-Guide for Agentic Engineering

2026-11-13 · Source: https://www.youtube.com/watch?v=Z-c11pV_uvU · Frank Coyle (UC Berkeley CS lecturer), talk on the AI Engineer channel, uploaded 8 Aug 2026, 20 min

Gist

Coyle uses Anthropic's Claude Certified Architect (CCA) exam as a map of what agentic engineering involves, whether or not you sit it. His claim: Anthropic sees how people use Claude, so the exam's scenarios show where the real problems are. He walks the exam's production scenarios and, for each, names the anti-pattern to avoid. Knowing what not to do is often how you get the question right.

A conference talk, mostly a tour. The code slides aren't in the transcript. Exam facts come from him and were checked only against third-party guides (see below).

Takeaways

• Philosophy. Opens and closes with Sister Corita Kent: "Nothing is a mistake. There's no win and no fail. There's only make." (attribution not checked). Experiment, build things, expect most of them to fail. Design patterns from 1990s OOP now have agent equivalents, and anti-patterns too.
• The exam, as he describes it. Released March 2026, scenario-based, timed, proctored, $99 for individuals, one try per 6 months. Five domains: agentic architecture 27%, Claude Code configuration and workflow 20%, prompt engineering and structured output, tool design and MCP integration, context management and reliability. Six production scenarios, four picked at random per sitting. Third-party guides agree on the March 2026 launch, six scenarios with four picked, and the 27/20 split. One guide notes the weights aren't published by Anthropic, and one lists the fee as $125. Not checked against Anthropic.
• Loops aren't new. Boris Cherny says his job is writing loops, not code. Peter Steinberger (OpenClaw) says he designs loops that prompt agents. Coyle: Böhm and Jacopini (1966) showed sequence, if-then, and loop are all a language needs to compute anything computable (date not checked here). Prompting had sequence and conditionals. Agents add the loop, which is where the power comes from.

Scenarios and anti-patterns

1. Customer support resolution agent. Anti-pattern: let the agent run and use whatever comes back. Instead, loop on the stop reason. The model can't run tools. It only proposes the call and its parameters, and your code runs it and feeds the result back. Loop while the stop reason is tool use. When it ends, check confidence and escalate to a human if low. Also check for running out of tokens, which means the answer is partial. He attributes stop reasons to Claude Code; they're a Messages API field (not checked here).
2. Code generation with Claude Code. CLAUDE.md files at three levels, top of project, project folder, and subdirectories, forming a hierarchy of rules. His description of the levels is muddled. Claude Code's docs (not checked here) name user, project, and subdirectory levels.
3. Multi-agent research system. Anti-pattern: one agent loaded with every tool. Analogy: you hire a carpenter and he shows up with plumbing and electrical tools too. Give each agent one job and one or two tools, like functions in functional programming. Keep subagent context out of the main context: tokens cost money, and more context confuses the model, even with a million-token window.
   • Critic subagent. Pass it only the claim and the evidence, not the reasoning that produced the claim. Agents that share everything drift into groupthink, like going along with pizza at a party. "Every agent gets its own slice."
4. Developer productivity. Subtask isolation, which he ties to multithreaded programming: threads that share memory need locks and synchronization, so keep threads, and agents, independent. Anti-pattern: every subtask dumps its full output into the main thread and context grows unbounded. Instead, fork the context for the subtask (example: scan logs for errors), return only a summary, and compact when the token count passes a limit (his example: 150,000). He isn't sure how compaction works inside.
   • Aside: a free book handed out at the conference, by a "Sam Bagwell" as transcribed, says his company sells custom context compression with an extendable base class. Possibly Sam Bhagwat (Mastra). Not verified.
5. Claude Code in CI. Anti-pattern: interactive mode in a pipeline, where Claude stops to ask permission. Configure it to run straight through.
   • Batch. Batched requests cost 50% less. He says results come in "at least 24 hours". Anthropic's docs say up to 24 hours, often sooner: https://docs.anthropic.com/en/docs/build-with-claude/message-batches
6. Structured data extraction. Listed in the agenda, not covered before time ran out.

Patterns

• learning what not to do points to what to do ← "anti-patterns are a key to understanding"
• a hyped idea is often an old primitive in a new medium ← "we are seeing this being resurrected"
• the model proposes actions, separate code executes them ← "the LLM is not executing these tools"
• check why a process stopped before trusting its output ← "check your stop reason"
• a specialist with few tools beats a generalist with all ← "specialize, don't overload"
• more context raises cost and lowers accuracy ← "context means tokens, tokens mean money"
• show a reviewer the claim and evidence, not the reasoning ← "a tendency to have group think"
• isolate side work, return only its summary to shared state ← "does not come back and pollute"