The map
The AI Engineering Skills Map is the spine of this deskAndrew Ng’s four skills — wrap, eval, ground, ship — are how this board is filed. If you are building an AI app, start here, not with a model card.
Latent Space5d agoEditorial
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.
The map
The AI Engineering Skills Map is the spine of this deskAndrew Ng’s four skills — wrap, eval, ground, ship — are how this board is filed. If you are building an AI app, start here, not with a model card.
Latent Space5d agoEditorial
Skipping permissions is the whole product question for coding agents. Read this as a security and UX harness, not a feature list.
The craft
Harness engineering is the new self-improvement storyLilian Weng’s map of recursive improvement is not sci-fi: it is loops, evals, and the wrapper around the model. If you only watch model cards, you will miss the job.
Lilian WengJul 4Editorial
Cookbooks for building AI apps: context, harness, graphs, cloud.
Wrap
Wrap GLM-5.3-Flash this week: 1M-context API, not a lab you train atCall it through Z.ai or Baseten — long-context wrap, MIT weights as a swap. Cookbook for the endpoint, not a training drop.
Z.ai4d agoEditorial
Context
Effective context engineering for AI agentsWindow, packing, routing: what belongs in context this turn. A lab you run on the wrap, not a model-card internals note.
Anthropic EngineeringEditorial
Graph
LangGraph persistence: checkpointed loops you can resumeDurable graphs, not a chatbot demo. This is the lab for threads that survive a crash in production.
LangChainEditorial
Eval
Trace the agent loop: OpenAI observability you ship withEval in the wrap: traces and integrations for agents you already call, not a training-lab dashboard.
OpenAIEditorial
App
Next.js App Router: a streaming AI app with the AI SDKBuild the chat UI and the route handler this week. Cloud lab for shipping an application, not wiring a network.
Vercel AI SDKEditorial
Wrap
Call Bedrock Converse in your app this weekWrap the AWS API — messages, tools, streaming — inside the product. Not a VPC lab.
AWSEditorial
Cloud
Deploy a Python ADK agent to Cloud Run this weekGCP lab: Vertex/ADK as an agent app on Cloud Run. Ship the product, not a TPU blog.
Google CloudEditorial
Cloud
Hosted agent on Azure AI Foundry: quickstartFoundry lab: stand up a hosted agent as an application in Azure, not a networking diagram.
Microsoft LearnEditorial
Cloud
MCP client on Cloudflare WorkersPut the harness on the edge: Workers talking MCP. A lab you take this week, not a training run.
CloudflareEditorial
Cloud
OpenAI Agents SDK in a Modal sandboxCloud lab: run the coding-agent harness in Modal, not on a training cluster. Wrap, eval, ship.
ModalEditorial
Harness
Effective harnesses for long-running agentsLoops, tools, permissions: the wrapper you ship around the model. Take this workshop before you write another agent from a model drop.
Anthropic EngineeringEditorial
Graph
Agentic RAG with LangGraph: retrieve inside the loopGraph lab you run this week: checkpointed retrieval, not a paper on embedding-model science.
LangChainEditorial
Cloud
WorkflowAgent on Vercel: a serverless harnessCloud lab: durable agent workflows on Vercel serverless, not a model-drop ticker.
Vercel AI SDKEditorial
MCP
Code execution with MCP: tools without stuffing the windowHarness lab: let the agent write code against MCP instead of packing every tool schema into context.
Anthropic EngineeringEditorial
Harness, loops, graphs, MCP, permissions, evals as acceptance.
Harness
Managed agents: decouple the brain from the handsAnthropic’s engineering note on splitting the planner from the tools is the production version of a harness. Same idea whether you like Claude or not.
Anthropic EngineeringEditorial
Clone-to-ship: RAG, eval harnesses, agent wrappers. Not train-a-model.
Repo
run-llama/llama_indexThe default RAG data framework: ingest, index, retrieve. Grounding before you wrap an agent.
GitHub9h ago52k ★Live
Repo
promptfoo/promptfoofor prompts and . Red-team and regression the wrapper, not the weights.
Watchlist posts for people building AI apps. Originals, not architecture notes.
X
No /dev/kvm, so the agent shipped a GitHub ActionHarness permissions are the product: the agent shipped a CI sandbox because it lacked /dev/kvm.
Simon Willison / X10d agoLive
X
Codex locked-use locked a Mac keychain — twicebrowser-use’s Mac harness is the app on the laptop in front of you: Accessibility, a computer-use loop, tool permissions.
browser-use/macos-harnessEditorial
Paper
WikiSkill: compile agent experience into persistent skillsA harness pattern, not a training run: turn traces into a wiki the next loop can actually reuse.
arXiv 2608.274542d agoEditorial
Paper
Persona vs execution: govern the hands, not the voicePersona-Execution Separation is the audit story for agents at work: a character in the window, a separate, logged loop that actually acts.
arXiv 2608.274272d agoEditorial
Weng’s 2023 LLM Powered Autonomous Agents post is how a generation learned planning, memory, and tools. Read it before the new jargon.
Lilian WengJun 23Editorial
If your agent cannot checkpoint, it is a demo. This is the open-source default for state, retries, and human-in-the-loop.
langchain-ai/langgraphEditorial
LangChain’s deepagents repo is what “just add an agent” looks like when someone actually ships the planning loop and the filesystem.
langchain-ai/deepagentsEditorial
Model Context Protocol is how you plug GitHub, a browser, or a database into many clients. Learn it as an interface, not a brand.
MCP Python SDKEditorial
Hugging Face’s barebones library is a philosophy: first-class code-acting agents, not a 40-file framework. Steal the idea even if you don’t vendor it.
huggingface/smolagentsEditorial
Agent Development Kit is how Google wants you to build, eval, and deploy agents. Worth knowing even if you never ship on Vertex.
google/adk-pythonEditorial
Official, lightweight, and opinionated about handoffs. Read it as a reference implementation of loops, not as the only way.
openai/openai-agents-pythonEditorial
A hosted harness, not a new model. The news is that “deep agent” is becoming a product category.
LangChain3d agoLive
Anthropic’s production MCP path: admins provision connectors through the identity provider. This is how you ship tools at a customer, not a demo OAuth loop.
Anthropic6d agoEditorial
GitHub18m ago25k ★Live
Repo
sandbaseai/deepseek-harness-handbookAgent-first DeepSeek Harness handbook: 173 source-backed runtime, plugin, MCP, sandbox, evaluation, troubleshooting, multilingual, and 74-resource Awesome ecos… Clone-to-ship: RAG, eval, or an agent wrapper.
GitHub3m ago91 ★Live
Repo
rojim666/SztuCode本地优先的 AI 编程 Agent，支持 TUI/桌面端、工具权限、会话记忆、Skills、Subagents 与 MCPA local-first AI coding agent with TUI and desktop clients, tool permissions, memory, Skills, Suba… Clone-to-ship: RAG, eval, or an agent wrapper.
GitHub4m ago45 ★Live
Repo
openai/codexOpenAI’s terminal coding agent. Watch it as a harness, not as a mascot.
GitHub4h ago120k ★Live
Repo
OpenHands/OpenHandsAn open coding agent with a real workspace. The other pole from a one-file CLI.
GitHub2h ago86k ★Live
Repo
letta-ai/lettaStateful agents with memory that lasts longer than a context window. Grounding that survives the next call.
GitHub6d ago24k ★Live
Repo
furkankly/zoetropeWatch a Claude Code session as a live flow graph. Read the harness, don’t guess what the agent just did.
GitHub5d ago663 ★Live
Repo
Aleph-Alpha-Research/eval-frameworkComprehensive LLM evaluation at scale: A production-ready framework for evaluating large language models across multiple benchmarks. Clone-to-ship: RAG, eval, or an agent wrapper.
GitHub34m ago41 ★Live
Repo
vercel/aiThe TypeScript toolkit if your loop lives in a Next.js app. Streaming and tools without a Python tax.
GitHub4h ago26k ★Live
Coding-agent sandbox vs OS keychain: a production harness footgun.
swyx / X4d agoLive
Lovable’s CTO on shaping the build: ship tools, not another consumer app. MCP is the interface.
Latent Space5d agoEditorial
Graph
Ng graph engineering: loops into working graphs (2h)Agentic knowledge graphs as app craft: first agent, how loops work, freeze them into graphs. Not a training lecture.
Anatoli Kopadze / X3d agoEditorial
Loop
Loops vs graphs: agent owns the path, or you draw the nodesLoop while the work is messy. Freeze it into a graph once the pipeline repeats. Same company brain either way.
Annatar.md / X6d agoEditorial
Loop
Karpathy: agents → loops → graphs as the app craftThe lecture is wrap/eval/ship: loops and graphs you run, not how a lab pretrains a model.
virgilxbt / X6d agoEditorial
Harness
Claude Code ELI5 skill: MCP and harness as a pageA skill that draws the wrapper: MCP runtime and harness, not a model card. Install it on a real ticket.
GoFly / X5d agoEditorial
Graph
Memory vs graphs: a pile vs a routeMemory is continuity. The graph is who may retrieve and write back. Routing, not a bigger embedding store.
Lunar / X5d agoEditorial
Graph
Ng’s agentic knowledge graphs course (1h)Persistent context for agents you ship: build the graph, then the multi-agent system. Not a grind thread.
Swati Gupta / X4d agoEditorial
Course
Hugging Face agents course: ship an agent, freesmolagents, LlamaIndex, LangGraph, then a project with evals. Catch up already files the lab; this is the pointer.
Ren / X3d agoEditorial
Harness
How to design an agent harness: work inside the system of recordAgents share data and policy with humans in one continuously updated view, with an audit trail. Harness design, not a valuation hook.
Yarchi / X3d agoEditorial
FDE
No-BS FDE guide: wrap and ship in the customer’s stackForward-deployed craft: you are in the account building the AI app. Not a salary-in-30-days thread.
Rahul / XAug 1Editorial
Claude
Full Claude Course: practical Claude for shipping an appWrap Claude, add tools, ship a real project. Practical course, not a hustle bait.
Jobescape / XAug 11Editorial
Learn how startups use GPT-5.6 to build faster, more cost-efficient AI agents with smarter model selection and new Responses API capabilities.
OpenAIAug 13Live
Models keep absorbing the harness into their weights — soon, it will be a harness for human attention rather than for the model.
Latent Space8d agoLive