Blog
- ai-agentsJul 14, 2026
Vercel Labs' Agent Browser: The Missing Feedback Loop for Coding Agents
Coding agents can write code, run it, even write tests — but they've never been able to check their own work in a browser. Agent Browser, a new Rust CLI from Vercel Labs, closes that loop.
- claude-codeJul 14, 2026
Claude Code Skills: Stop Repeating Yourself, Stop Bloating CLAUDE.md
A Skill is a folder with a SKILL.md file — a reusable workflow Claude discovers on its own based on the task, instead of you retyping the same instructions every session.
- ai-engineeringJul 8, 2026
Headroom: Context Compression Before the LLM
You build an agent, send 20,500 tokens, and the model may only need 6,700. Headroom shows why context compression before the LLM becomes its own infrastructure layer.
- ai-engineeringJun 10, 2026
Building Production AI Systems
Why engineering fundamentals — not prompt tricks — decide whether your AI system survives contact with real users.
Claude Code Deep Dives
Claude Code Rules: Stop Dumping Everything Into One Giant CLAUDE.md
A 500-line CLAUDE.md gets loaded in full, every session, whether the task needs it or not. Rules split it into small, path-scoped files that load only on demand.
Claude Code Security Patterns: The Permission Rules I Wish I'd Set Up Sooner
Clicking 'don't ask again' only covers that one exact command. Here's the permission system that actually scales — and the settings hierarchy that decides which rule wins.
Claude Code Skills: Stop Repeating Yourself, Stop Bloating CLAUDE.md
A Skill is a folder with a SKILL.md file — a reusable workflow Claude discovers on its own based on the task, instead of you retyping the same instructions every session.
Claude Code Agent Teams: When Sub-Agents Aren't Enough
A sub-agent reports back once and disappears. An Agent Team stays in the room — teammates share a task list, message each other, and challenge each other's findings before you see the result.
Claude Code Routines: Automate Recurring Work Without Sitting at Your Laptop
A Routine is a Claude Code job — trigger, self-contained prompt, connectors — that runs on Anthropic's managed servers. No laptop required. Here's how it actually works.
More articles
Vercel Labs' Agent Browser: The Missing Feedback Loop for Coding Agents
Coding agents can write code, run it, even write tests — but they've never been able to check their own work in a browser. Agent Browser, a new Rust CLI from Vercel Labs, closes that loop.
Monitoring AI Agents Locally: Langfuse + OpenTelemetry, No Cloud Account Required
Your agent works in a demo. Then it ships, something breaks in production, and you have zero visibility into what it actually did. Here's how to trace every run — self-hosted, on your own machine.
Context7: Fixing the One Thing Every AI Coding Assistant Gets Wrong
Your LLM's training data is a snapshot from the day training stopped. Context7 fetches the actual current docs for the library you're using, right before the model writes the answer.
Headroom: Context Compression Before the LLM
You build an agent, send 20,500 tokens, and the model may only need 6,700. Headroom shows why context compression before the LLM becomes its own infrastructure layer.
Building Production AI Systems
Why engineering fundamentals — not prompt tricks — decide whether your AI system survives contact with real users.
Evals That Actually Work
Most "evals" are vibes in a trench coat. Here's the smallest harness that gives you a signal you can trust.
ChatGPT Apps SDK: Build an App That Runs Inside ChatGPT
Apps inside ChatGPT aren't another chatbot wrapper. You ship a product surface where users already talk to the model — architecture, local Todo app, ngrok, then wire it up.
Google's Gemini CLI: gemini.md, MCP, and a Real Next.js Bug
A new AI CLI is useless until it can fix something in your repo. Gemini CLI got interesting for me once gemini.md, MCP, and a real bug were in the same loop.
llms.txt: Stop Letting Coding Agents Invent Last Year's API
The quiet failure mode in AI coding isn't wild hallucinations — it's confident use of an API that renamed itself after the cutoff. llms.txt + mcpdoc feed current docs into the agent.
Transcribe Any YouTube Video with Gemini 2.5 Pro
A tiny uv + Gemini script: pass a YouTube URL, get a clean summary. No manual download — use it as a filter before you invest watch time, not as a replacement for accuracy-critical viewing.
browser-use: Web Automation Agents Without the Selector Tax
Classic browser automation dies on brittle CSS selectors. browser-use flips it: natural-language goals, model-driven clicks, then a real LinkedIn automation path from empty env to working agent.
AI engineering, weekly.
Join developers getting practical AI engineering in their inbox.