Setting type
Agentic Firehose
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.
Curriculum · four skills, not a feed
The job, in recruiter language — then Ng’s four skills. This is not a fifth skill. Came for harness / loop / graph? Decoder first. Company repo this week? Starting a project.
This is the job
Not a fifth skill
Orchestrate models into systems
01 Make an AI thing that actually worksBuilding and deploying AI applications
Architecture judgment
02 Software engineering fundamentalsTradeoffs, tests, security — so you can steer the tools
Not just AI assistance
03 Using coding agentsWhen to grab the wheel, and how to check the work
Decide what to wrap
04 Shaping the buildProduct sense, specs, MVP vs a careful build
The four-skill spine — not a fifth skill
Same 01–04. Plain English first, Ng’s name under it. Nested items are DeepLearning.AI and similar — start at 01 unless you already have the floor. The map.
01
Building and deploying AI applications
Ng’s first skill. The six tags on the board live here: LLMs / context, Grounding / RAG, Agents / harness, Evals, Production, ML foundations.
1.01One letter
Andrew Ng’s four-skill map is the spine of this page. Read it once so the firehose has a filing cabinet.
DeepLearning.AI / The Batch
1.02One letter
The follow-up letter is the vocabulary on every headline. Unpredictable outputs, then evals that make them governable.
DeepLearning.AI / The Batch
1.03A sitting
Specify, iterate, eval. Prompts are the interface, not the job. Start here for LLMs / context.
DeepLearning.AI
1.04Short course
RAG was the first trick. This short course is the vector-search version of Grounding / RAG before you reach for a graph or tools.
DeepLearning.AI
1.05Short course
Chained calls, evals, and the software around the model. The turn from a clever prompt to an application.
DeepLearning.AI
1.06Short course
Quality and Safety for LLM Applications is the disciplined eval / error-analysis loop Ng says separates people who ship.
DeepLearning.AI
1.07~10 hours
Ng’s vendor-neutral course: reflection, tool use, planning, multi-agent — then evals. This is Agents / harness without a framework religion.
DeepLearning.AI / Andrew Ng
1.08Short course
State, nodes, edges, a loop you can debug. The curriculum version of the systems lane.
DeepLearning.AI / LangChain
1.09One essay
Anthropic’s Building Effective Agents, read as craft not brand. Workflows first; agents when you need them to steer.
Anthropic Engineering
1.10Long essay
Lilian Weng’s map of loops, evals, and the wrapper. After this, the splash stories stop sounding like sci-fi.
Lilian Weng
1.11Weeks, not days
Ng’s Machine Learning Specialization: training, overfitting, error analysis. You do not need to become a researcher.
DeepLearning.AI / Andrew Ng
1.12Weeks
The Deep Learning Specialization is the other half of ML foundations: nets, sequence models, and why transformers exist.
DeepLearning.AI / Andrew Ng
1.13Monday stack
Classify, score, anomaly — in the app that already runs the company. The wrap is the product. Not a training run.
Starting a project · 07
02
Software engineering fundamentals
Tradeoffs, architecture, testing, security. The precise language you use to steer a coding agent — instead of vibe-coding the wrong stack.
2.01Short course
If you cannot write a script, you cannot read what the agent just did. Ng’s AI Python for Beginners is the on-ramp.
DeepLearning.AI
2.02A repo
smolagents is a small, readable harness: the model thinks in code. Clone the loop — not a train-a-ChatGPT stack.
huggingface/smolagents
2.03A repo
MCP is how tools plug in without marrying a vendor. Learn it as an API design problem — the software-engineering half of a harness.
MCP Python SDK
2.04A repo
LangGraph’s durable graphs are the architecture lesson: checkpoints, retries, human-in-the-loop. Name the tradeoff before the agent picks one.
langchain-ai/langgraph
2.05One essay
Willison’s prompt-injection notes are the security module. You need this language to tell a coding agent what not to touch.
Simon Willison
2.06Monday stack
SSO, tickets, the old app. The agent has to live where the company already works. Skip only if this is greenfield.
Starting a project · 07
03
Using coding agents
Context, when to intervene, verifiers and evals, multi-agent coding, and a habit of evolving the workflow as the tools move.
3.01One letter
The skills-map letter: mental model, how much to intervene, verifiers so the agent can close its own loop, pitfalls like the production database.
DeepLearning.AI / The Batch
3.02A repo
Karpathy’s autoresearch: program.md is yours; train.py is the agent’s. A verifier (the metric) is how the loop closes without you.
karpathy/autoresearch
3.03A thread
The thread behind autoresearch. Most “agent products” still have this split backwards.
Karpathy / X
3.04A repo
OpenAI’s Codex CLI is one current shape of the skill. Watch how it uses context and tests, not the mascot.
openai/codex
3.05A repo
OpenHands is the other pole from a one-file CLI: files, commands, when to grab the wheel. Practice intervening.
OpenHands
3.06Self-paced
Hugging Face’s Agents course is free and not married to one vendor. Good complement when you want to see the same loops in Python.
Hugging Face
04
Shaping the build
Product sense, specs, MVP vs a careful build, ownership. Agents are getting better at implementing a spec — someone still has to decide what is in it.
4.01One letter
Ng’s fourth skill: pixel-perfect handoff is dying. You need business context and customer goals to shape what gets built.
DeepLearning.AI / The Batch
4.02A few hours
AI for Everyone is still the cleanest picture of what this field is for, without pretending you already train models.
DeepLearning.AI / Andrew Ng
4.03Ongoing
Latent Space interviews are how you borrow product sense from the people writing harnesses — not the people announcing apps.
Latent Space
4.04Weekly
The Batch is Ng’s weekly letter. No public RSS, so you subscribe. It is the catch-up layer after you have a path — including updates to this skills map.
DeepLearning.AI / The Batch
This is a curriculum, not a feed. Each skill has Ng’s name and a plain-English line so a first visit does not require the glossary. Skip a skill only if you can already do the work it names. The six tags on the board are the sub-skills under “Building and deploying.” When a nested item is done, go back to today’s news — it will file itself. Ng’s skills map.