From the lab notebook.
Engineering notes from shipping AI in production.
Kimi K3: Moonshot AI's 2.8T open-weight model closes the gap
Moonshot AI's Kimi K3 debuted at #3 on the Artificial Analysis leaderboard, beat Opus 4.8 and GPT-5.6 Sol on agentic coding benchmarks, and is shipping as open weights — with a price tag to match.
Alibaba's Page Agent: a browser agent that skips screenshots entirely
Alibaba open-sourced Page Agent, a JavaScript library that controls web pages through the DOM instead of vision models — no headless browser, no screenshots.
Alibaba Zvec: the new embedded vector database for on-device RAG
Alibaba’s open-source Zvec makes vector search local, fast, and durable with a SQLite-like in-process experience.
How to Pick the Right LLM for Your Agent
Not every agent needs the most capable model. Here's a practical framework for matching model to task — and why the answer almost always changes over time.
MCP Explained: The Protocol That's Quietly Changing AI Tooling
Model Context Protocol (MCP) is becoming the USB-C of AI integrations. Here's what it is, why it matters, and what it means for how you build agents.
Why Your AI Agent Should Be Model-Agnostic
Locking your agent to a single model provider is the fastest way to accumulate technical debt. Here's how to build the abstraction right.
What Is an AI Agent, Actually?
Everyone is shipping 'agents' — but most are just chatbots with a system prompt. Here's the real distinction and why it matters.
Evals are the product. Everything else is a side effect.
After two years shipping LLM features in production, the only artifact worth trusting is a well-designed eval suite. Here is how Cuecoder structures them.
Stop tuning prompts. Start tuning context.
Prompt engineering plateaus fast. The real lever is what you put in front of the model — retrieval, structured tools, and dynamic memory.
Building agents without frameworks
LangGraph, Crew, AutoGen — all good demos, none have shipped what production needed. Here's the 200-line loop quietly serving production traffic.
The 1M context window is a trap (for now)
Bigger context windows are exciting and useless without a retrieval strategy. A pragmatic guide to picking what to put in.