Setting type
Claude Computing
Pulling labs, repos, and the catch-up path…
Harnesses, graphs, loops, APIs you call this week — tagged with Ng’s application skills. Not a training desk. New here is the decoder. Starting a project this week? Catch up is the four-skill map — not more news.
Heard it on X. Here’s the decoder.
Context, harness, loop, graph — these are not competing schools. A model needs something to see, hands to act, a repeat-until-done, an org chart when work branches, a test to know it worked, then you ship it. Layers of one job. Starting a project is the Monday install order if you already have a company repo. The mission is architecture judgment plus AI orchestration.
How it fits together
what the model can see this turnboard tag LLMs / context
You choose what goes in the window — files, tool traces, retrieved facts — instead of pasting the whole repo and hoping.
Start with Talk to models on purpose — DeepLearning.AI, A sitting. Catch up 01
fetch the facts, don’t guessboard tag Grounding / RAG
Look up the right documents at ask-time and feed them in. A lookup plus a write-up — not an agent by itself.
Start with Ground the model with data — DeepLearning.AI, Short course. Catch up 01
the wrapper around the modelboard tag Agents / harness
Tools, permissions, and checks so the model can finish a job. The brain is in a jar; the harness is the hands.
MCP is the USB port for those tools — Catch up skill 2.
Start with Agentic AI, from first principles — DeepLearning.AI / Andrew Ng, ~10 hours. Catch up 01
look, act, check, repeatboard tag Agents / harness
The smallest useful agent is not a chatbot. Read the result, do the next thing, check again, until the job is done.
Start with Systems, not one-shot prompts — DeepLearning.AI, Short course. Catch up 01
branches, specialists, handoffsboard tag Agents / harness
When one loop isn’t enough: who runs next, when to retry, when to ask a human. An org chart, not a circle.
Start with Agents as graphs, not vibes — DeepLearning.AI / LangChain, Short course. Catch up 01
did it actually finish the jobboard tag Evals
A test the system can fail. Without it you cannot tell if the harness got better or just sounded better.
Start with Evals you can actually run — DeepLearning.AI, Short course. Catch up 01
ship it so it stays upboard tag Production
Cost, latency, crash-and-resume, guardrails. A loop that only works in a notebook is still a demo.
Start with If it cannot crash and resume, it is a demo — langchain-ai/langgraph, A repo. Catch up 02
Leave with a plan
Not a new catalog. Same DeepLearning.AI path as Catch up skill 1, then skill 2.
This sitting
Talk to models on purposeContext engineering in one sitting — specify, iterate, eval.
DeepLearning.AI · A sitting · Catch up 01
Start this week
Agentic AI, from first principlesNg’s Agentic AI course is harness + loop without a framework religion.
DeepLearning.AI / Andrew Ng · ~10 hours · Catch up 01
Next
Agents as graphs, not vibesGraph engineering: state, branches, a loop you can debug.
DeepLearning.AI / LangChain · Short course · Catch up 01
Then
Evals you can actually runEvals you can actually run, so the loop is not vibes.
DeepLearning.AI · Short course · Catch up 01
Then Catch up 02
Python that actually shipsSoftware fundamentals — the language you use to steer coding agents.
DeepLearning.AI · Short course · Catch up 02
After that, stay on Catch up— Ng’s four skills are the spine. Then come back to today’s news; the chips will make sense.