Anthropic senior engineer just dropped a 1-hour course on building self-learning agentic systems with loops & graphs: • 00:19 - how Claude Code actually works • 14:20 - long horizon agents CLAUDE.md & Plan mode • 23:38 - Claude "skills" design patterns for agents • 44:01 - building agents... & subagents with Claude • 59:34 - self-improving loops & graphs for Claude agents This 1-hour watch will replace a $500 agentic engineering course Watch today, then read how to build self-improving agents with "loops" and "graphs" in the article belowshow more

@0xCodez digging into the Claude "skills" section opened my eyes. I hit a wall myself on agent design until I reconsidered skill planning.

amazing workshop with Claude Code team !

yeah, one of the best watch on building agents with Claude in my opinion !

The part after the loop deserves a timestamp too: who checks the result, and what earns another iteration? Our Agentic Adderall stack puts adversarial review inside the run. Otherwise you have invented a very confident while(true).

This video is from April ?

legendary watch, i'm building the same loops right now. thx bro

the 14:20 section is the one i’d watch first, because CLAUDE.md and plan mode encode persistent guidance without pretending the model has durable memory. does the course show how those instructions are kept from becoming stale or contradictory?

is there a youtube link for this?

The loop matters less than who validates it. Same model proposing and marking it passed converges on its own blind spots. We split that into three roles with no shared context — find the issue, fix it, accept it — because self-review confirms, it doesn't catch.

plan mode + skills is the real unlock

an hour long course from a senior anthropic eng covering how Claude Code actually works under the hood — long horizon agents and skills design patterns are exactly the parts the docs skip. saving this one.

The long-horizon agents section is the one to watch.

Hi, I'm here today to teach you how can you use a huge amount of tokens.

A course is useful when it changes the system you ship, not when it adds another hour to your watch list.

great explanation about loop engineering

The 'latest' best model in the video is Opus 4.7 - so this video must be from back in April? A fair bit has changed since then. But still useful