Anthropic released a Claude Certified Architect exam in March, and @coyle_frankp treats it as something more useful than a credential: a map of what Anthropic already knows about how agent systems break. His talk "Anthropic's CCA Exam as a Field-Guide for Agentic Engineering" is on @aiDotEngineer's YouTube. Frank lectures at UC Berkeley after 32 years teaching computer science, and he came at this looking for a way to get his students ready for agentic work. The value here isn't exam prep. It's that he walks the exam's six production scenarios and, for each one, names the anti-pattern, on the argument that knowing what not to do is what leads you to what you should. - The exam's shape. $99 for individuals, once every six months, timed and proctored, scenario-based multiple choice with realistic constraints. Five domains: agentic architecture at 27%, Claude Code configuration at 20%, prompt engineering and structured output, tool design and MCP, context management and reliability. - Loops aren't new, they're Turing completeness arriving in agents. Böhm and Jacopini proved in 1966 that sequence, conditionals, and a loop are all you need to compute anything. The agentic world just rediscovered the third one. - Read the stop reason, don't just take the response. The LLM can't execute a tool. It stops and hands you parameters. Loop on the stop reason, run the tool, feed the result back. A stop reason can also mean you ran out of tokens and the answer you're holding is partial. - Don't hand one agent every tool. You hire a carpenter and he shows up with plumbing and electrical gear too. Maybe you want the carpenter. One or two tools per agent. - Starve your sub-agents of the reasoning that produced the input. Give a critic agent the claim and the evidence, not the thought process behind them. Agents that see each other's reasoning converge into groupthink the way a room talks one person into pizza. - Fork context for subtasks, then merge only the summary. Scan the logs in a forked context, return the summary, keep the noise out of the main thread. Check your token count and run a compaction when it crosses your threshold. - CI runs need non-interactive mode. An interactive Claude will stop in your pipeline and ask for permission. Also worth knowing: batch mode is 50% cheaper if you can wait 24 hours for results. I'm working through the published talks from AI Engineer World's Fair sharing summaries and takeaways. Follow for more!