Karpathy's 630-line Python script automates 50 experiments overnight

🐍 10 Python projects that actually make you better (not just busy). Stop getting stuck in tutorial hell. If you want to actually level up your Python skills this year, you need to build things that solve real problems. Here are 10 projects that will teach you core engineering concepts: https://lnkd.in/ge_-4Hby https://lnkd.in/g9HDS32J

Over the past few days, I built another Python library: 𝗔𝗴𝗲𝗻𝘁𝗞𝘂𝗯𝗲-𝗠𝗶𝗻𝗶. A tiny agent orchestration engine.  ~239 lines of Python, zero dependencies It started with a simple question: What is the minimum needed to orchestrate multiple agents? Not frameworks. Not abstractions. Just the core. So I stripped it down to: Task DAG → agents + dependencies Scheduler → runs independent tasks in parallel Event system → observe every step Shared memory → pass outputs downstream If you look at most multi-agent systems, they follow the same pattern: nodes (agents), edges (dependencies), executed by a scheduler Everything else is layering. It’s for a different purpose: → to understand orchestration by reading it end-to-end Try: `𝙥𝙞𝙥 𝙞𝙣𝙨𝙩𝙖𝙡𝙡 𝙖𝙜𝙚𝙣𝙩𝙠𝙪𝙗𝙚-𝙢𝙞𝙣𝙞` Because once you see it in ~200 lines, a lot of the “magic” disappears and that’s a good thing. Curious, how are you orchestrating your multi-agent workflows today?

If you use Cursor or Claude Code to write Python you've probably noticed LLMs have some odd habits. Claude especially loves hasattr/getattr calls, unexplained function-scope imports, bare excepts, passing True/False as positional args with no context, and defaulting to dataclasses when your project uses Pydantic. I use agents pretty heavily to build MA-Gym (my open source multi-agent RL library, v-1.0.0 pending soon) and a lot of our internal tooling at DeepFlow. I kept catching the same stuff in review, so I wrote a linter in Rust to do it automatically. slopcop. 16 rules, tree-sitter based, zero false positives on strings/comments, sub 100ms, drops into CI/CLI in one line. pip install slopcop https://lnkd.in/eiUK2j4p Tell me what stupid Python idioms your agents keep writing and I'll add rules for them.

LangChain Revision #4: Prompts in Langchain Why Use Prompt Templates Over Python f-strings? As Python devs, f-strings feel natural for handling placeholders. So why does LangChain push you toward PromptTemplate instead? Turns out there are 3 solid reasons: 1. Validation — automatically checks if all placeholders are filled, catching errors during development rather than at runtime on a production server 2. Reusability — templates can be saved as JSON files and loaded across multiple apps and pages, keeping your code clean and modular 3. Ecosystem Integration — designed to work directly with LangChain Chains, so you can tie a template and a model together and invoke the entire pipeline in one call That's the kind of thing f-strings simply can't do. Drop a comment if you find it useful! Follow + 🔔 to get notified as the series grows. Revision #4 | LangChain Revision Series hashtag#LangChain hashtag#GenAI hashtag#LLMEngineering hashtag#RAG hashtag#GenerativeAI

Most data engines serialize data across process boundaries every time a Python User-Defined Function runs. Think Spark shipping rows to a separate Python process and back. Long ago, we eliminated that boundary entirely in Sail. Rust + PyO3 + Arrow, same process, zero copy.

Julia vs Python isn’t a technical debate. It’s a business one. Python wins on ecosystem and hiring. Julia wins on performance and elegance. In enterprise: - ecosystem usually wins - maintainability beats purity Context decides. Takeaway: Pick what reduces total system friction. I've seen Julia annihilate Python from a performance perspective, yet I've never released a Julia production app. Have you seen Julia outperform Python in production?

Experimented with OpenClaw using local Ollama but performance was horrible (each response taking up 1.5+ minutes), so ended up writing a Python script for the task I had in mind for it instead. Uses Google OAuth to get access to Gmail, goes through recent messages and marks anything that doesn't look important as read. https://lnkd.in/gAHcxcdv #MinorInconveniencePatches : #3

Is OpenAI buying python and Anthropic buying typescript? Uv and bun bought respectively. Hmmmmm Codex for Python, Claude Code for Typescript? Obviously both can do both, but I wonder if this might be a drawing of lines in the sand?

Litecrew Launches Minimalist Multi-Agent Orchestration for Python Developers 📌 Litecrew just dropped a minimalist Python tool that lets devs orchestrate AI agents in under 150 lines - slashing hours off prototyping. No heavy frameworks, no YAML nightmares: just clean code, simple patterns, and fast results. Perfect for testing multi-agent workflows before scaling up. 🔗 Read more: https://lnkd.in/daJjZ9m7 #Litecrew #Python #Multiagent #Sqlite #Orchestration

I’ve wanted to write something in Rust for a while, and recently got a bit excited about the idea of a Python library in Rust. So I open-sourced fast-ordset - an ordered set for Python implemented in Rust (PyO3 + indexmap). In benchmarks it’s about 2–10× faster than the pure-Python ordered-set on various operations. The extension is thread-safe and doesn’t hold the GIL, so it works with free-threaded Python. One downside: iteration is noticeably slower - each element crosses the Rust→Python boundary (serialization/copying), so for x in s and list(s) are slower than the Python implementation. For indexing s[i] and bulk operations it’s fine. Install: uv sync --all-extras and uv run maturin develop --release. MIT license. Repo: https://lnkd.in/dXRXdWqY #Python #Rust #OpenSource