🤖 China Open-Sourced the AI Agent Playbook

A 14,183-star open-source book on building AI agents dropped to trending today. It is in Chinese. It does not matter.

The author, Li Bojie, shipped the full text, a compiled PDF, and chapter-by-chapter code in one public repo. Architectural theory, engineering practice, worked examples. The kind of thing that used to cost $400 at a conference and arrived three years after the ideas went stale.

Here is what I read into it: the agent engineering canon is no longer gated behind English-language labs and Western conference circuits. The practitioners who built the infrastructure those labs run on just wrote the field manual and handed it to everyone at once.

Meanwhile, Kimi K3 from Moonshot had to pause new subscriptions this weekend because demand overwhelmed capacity. That is not a PR story. That is a signal about where the weight of the developer community is shifting.

The shortlist today has 9 repos and a few infrastructure moves worth your time. Agent orchestration, fine-tuning, local inference, and a curated sub-agent prompt library that belongs in your Claude Code workflow. Let's get into it.

New here? This part's for you. (Operators, skip to The Drops.)

New here? This section is for you.

### What is open source?

Most software is locked. You can use it, but you cannot see how it works, copy it, or change it. Open-source software is the opposite: the code is public, the license says you can use it freely, and anyone can build on top of it.

This matters for AI builders because the best tools right now are open source. Someone builds a thing, publishes the code on GitHub (a website where developers store and share code), and thousands of people use it, improve it, and build new things on top of it. No purchase required.

### Do this today

Find today's top Drop below: the AI agent book with 14,183 stars. Click the link. You do not need to read Chinese. Scroll to the /code folder. Look at the structure of one chapter's example. You are not installing anything. You are learning how a working operator organized a real agent project.

### How to use today's picks

Today's issue has 9 repos plus a Stack tool worth knowing. For each one: read the one-line pitch, click through, and skim the README (the project's front-page notes on what it does and how to start). You are not committing to install anything. You are building a mental map of what exists. When a problem shows up in your work, you will know exactly where to go.

### Plain English

Repo, short for repository. A folder on GitHub holding a project's code, instructions, and history. When I say "clone a repo," I mean download a copy to your own computer.

Stars, on GitHub, a star is roughly a bookmark. High star counts mean many developers found the project useful enough to save. It is not a perfect signal, but 10k+ stars on a new repo means the community showed up fast.

### We read your replies

You told us you are here to learn, so we built this section. Hit reply and send us one word or phrase you keep seeing in AI content that you want decoded. We read every one.

[Repo] bojieli/ai-agent-book, the full-text open-source book on AI agent design and engineering practice, by Li Bojie, with chapter code included. 14,183 stars. This is the most complete operator-readable treatment of agent architecture I have seen shipped as a free public resource. If you build agents, read this even if you run it through a translation layer.

[Repo] HKUDS/DeepTutor, a lifelong personalized tutoring agent that adapts to what a learner already knows. 28,816 stars and it is still climbing. The architecture behind adaptive knowledge-state tracking is directly applicable to any agent that needs to avoid re-explaining context it has already covered.

[Repo] ayush-that/sub-agents.directory, a curated library of 100+ sub-agent prompts and MCP servers for Claude Code. 135 stars, but this is the most on-audience repo in today's shortlist. If you run Claude Code daily, this is a prompt library you can drop in today. No install required.

[Repo] aipoch/medical-research-skills, hundreds of agent skills covering protocol design, data analysis, evidence synthesis, and academic writing for medical research. 1,498 stars. The skills structure here transfers cleanly to any research or document-processing agent, not just medical. Worth reading the architecture before copying the domain.

[Repo] dzhng/skills, a structured skills library for agent composition. 622 stars. The shortlist snippet is sparse, but the author (David Zhang) has a track record with agentic tooling worth the click.

[Repo] pingdotgg/t3code, T3 Chat's open-source stack from Ping.gg. 14,236 stars. One of the fastest-growing repos in the shortlist today. If you are building a chat-layer on top of model APIs, the architecture decisions here are worth an hour of your time.

[Repo] mudler/LocalAI, run any model locally: LLMs, vision, voice, image, video, no GPU required. 47,728 stars. The open-source engine for operators who want inference on their own hardware without the API bill. If you have not run it against your current stack, that is the experiment for this week.

[Repo] axolotl-ai-cloud/axolotl, a fine-tuning framework that handles the config complexity so you can focus on the dataset. 12,227 stars. The go-to tool once you have decided a base model needs to specialize. Supports most modern fine-tuning patterns.

[Repo] huggingface/peft, state-of-the-art parameter-efficient fine-tuning from Hugging Face. 21,427 stars. LoRA, QLoRA, and the other methods that let you adapt a large model on a single consumer GPU. Pair it with Axolotl above and you have a complete fine-tuning workflow.

[Affiliate] Runpod, Axolotl and PEFT above get you fine-tuning, but not the hardware to run it on. Runpod rents H100 and A100 GPUs by the second, or deploys a model as a serverless endpoint that scales to zero, so you pay only while it runs. We may earn a commission.

[Repo] krokozyab/Agent-Fusion, a local RAG semantic search engine that gives your agents instant access to your codebase, Markdown docs, Word files, and PDFs. 74 stars, new to trending. The key thing here is local: your code and docs never leave your machine. Query your entire repo from inside a Claude Code session without pushing anything to a third-party embedding service. Set the watched directories once; it indexes on change.

[Repo] levi-qiao/graphkit, runs long coding tasks as a graph of agent nodes with an executor and a clean-context supervisor, rather than one drifting loop. 27 stars but the framing is exactly right. Loop drift is the failure mode that kills long-horizon agent tasks. A graph structure with a dedicated supervisor node addresses the root cause, not the symptom. Worth watching as it matures.

- Nativ lets you run AI models locally on a Mac with no terminal required. Prince Canuma, the developer behind MLX-VLM, shipped this as a native Mac app. If your operator workflow involves recommending local inference to non-technical collaborators, this removes the biggest barrier: the command line. (Simon Willison)

- Kimi K3 paused new subscriptions after demand overwhelmed capacity. Moonshot's model drew crowds at the World AI Conference and then ran out of room. The headline is demand; the signal is that the operator community is actively testing Chinese open-weight models at scale. The cost floor keeps dropping and it is not coming from one direction. (Euronews)

- Substack is adding an AI content detector powered by Pangram. It scans posts, notes, replies, and comments for AI-written content. For operators who publish on Substack or compete with Substack writers: this changes the surface-area calculation. A tool that flags AI-assisted content as a negative signal is a platform-level policy decision, not just a feature. (Engadget)

- Vercel Connect now ships with 90+ preset connectors including Shopify, Jira, Okta, and Workday. Preset configurations mean no manual setup for supported services. If you deploy agent pipelines on Vercel that need to touch enterprise data sources, the integration friction just dropped. (Vercel)

- Microsoft is facing new legal claims tied to alleged overstatements about AI product capabilities. The pattern matters: bold AI marketing claims are becoming a litigation surface. If you ship agent products with performance guarantees, read how Microsoft's claims are being characterized. (DataCenterDynamics)

Move: Wire a local RAG index to your Claude Code sessions for instant codebase Q&A.

1. Clone Agent-Fusion and point it at your project directory and any Markdown docs alongside it. 2. Let it index. Then open a Claude Code session and query it with natural language: "What does the auth module do?" or "Which file handles the webhook parser?" 3. Use the output as grounding context before you ask Claude to modify anything. Paste the relevant snippet as system context, not as a vague instruction.

You will know it worked when Claude stops hallucinating file paths and starts citing the actual function names in your repo. The difference between a session that drifts and one that stays accurate is usually whether the model has real retrieval behind it or is pattern-matching from memory.

No embedding API. No data leaving your machine. Fifteen minutes to first query.

We build The AIgent's engine in the open. An honest look at what we are making, what broke, and where it is headed. This week: a row we almost wrote into production that the system would have ignored forever.

A default is a decision somebody else already made for you, and it is silent by design.

We were seeding new rows into a content table. The insert was clean, the columns lined up, the data was right. It would also have been invisible. The table's status column defaults to a staged value, and the code that ships content only ever reads rows marked live. A row inserted without explicitly setting that flag lands in the table looking present, correct, and complete, and never appears in a single issue. It is not missing. It is not broken. It simply is not seen.

That is the dangerous shape of a silent default: nothing errors. A missing row throws. A malformed row throws. A row that is merely inert throws nothing at all, so you find it weeks later, if you find it.

What saved the seed was reading the consumer, not the schema. The schema tells you what a table will accept. It does not tell you what the rest of the system will actually use. Before you write to any store with defaults, find the query that reads it, and name every column that gates whether your work is used rather than merely stored.

Present is not the same as active. The gap between them is where quiet work goes to die.

What is sitting in your database, correct and completely ignored? Hit reply. We read every one.

The AI agent book trending at 14k stars is the kind of artifact that would have cost a developer conference badge and a cross-country flight two years ago. Today it is a free public repo. That compression keeps happening. The operators who notice it first are the ones who show up next week with something built.

Before you go: we started a room for people actually building with agents. Post what you shipped, get unstuck, and tell us what to build next. Join the community →

See you Thursday.