Search the archive

CURSOR CEO BUILT A BROWSER USING AI, BUT DOES IT REALLY WORK?

Cursor's CEO used hundreds of AI agents to build a web browser in a week, but the resulting software only "kind of" works and is far from production-ready.

THE AGENTIC AI HANDBOOK: PRODUCTION-READY PATTERNS

This is a collection of 113 production-ready patterns derived from real-world systems for building and deploying reliable AI agents.

SERVING MARKDOWN FOR AI AGENTS

Serving Markdown via GitHub Actions and Jekyll improves data ingestion over messy HTML.

REQUIEM FOR ANOTHER HELLSITE

Stack Overflow question volume collapsed from 250,000 monthly posts to 2009 levels by 2026, mostly due to AI.

* Unlimited parallel test runs

* 24-hour maintenance and on-demand test creation * Zero flakes, guaranteed

Schedule a demo to learn more

SCALING LONG-RUNNING AUTONOMOUS CODING

Single agents struggle with complex projects, and initial attempts at dynamic multi-agent coordination fail due to bottlenecks and risk-averse behavior. The Cursor team developed a successful "Planners and Workers" model, where Planners create tasks and Workers execute them independently, solving coordination issues. This system allowed hundreds of agents to run concurrently for weeks, generating over a million lines of code for ambitious projects like building a web browser and performing large-scale code migrations.

WE ROLLED OUR OWN DOCUMENTATION SITE

Tangled decided to build its own documentation site using Pandoc to meet specific requirements like no JavaScript, low complexity, and easy styling. Its team used Pandoc's "chunkedhtml" output to convert a single Markdown file into a static site, using an autogenerated table of contents and a JavaScript-free collapsible sidebar via the new popover API. For search, they used a Google redirect and provided a single-page version for browser-native Ctrl+F functionality.

I HATE GITHUB ACTIONS WITH PASSION

This dev hates GitHub Actions, mostly due to a frustrating incident where a `build.rs` script relying on the CUE binary failed on a Linux ARM runner. The issue happened because GitHub Actions' isolated environments prevented the CUE binary installed for x86_64 from being accessible to the ARM runner, leading to a very tedious and slow debugging loop.

THE TRAPS OF BORROWED INTERNAL DEVELOPER PLATFORM ROADMAP CERTAINTY

Internal developer platform roadmaps accumulate debt by treating borrowed industry trends as certainties. To avoid building useless features, teams must treat roadmap items as testable hypotheses.

DISCARDING THE SHAFT-AND-BELT MODEL OF SOFTWARE DEVELOPMENT

AI coding agents like Claude Code will shift software from bloated mega-projects to bespoke, artisanal software. By driving small-project costs toward zero, vibe coding allows for personalized applications for the 50 percent of engineers building internal tools.

TOOL SEARCH NOW IN CLAUDE CODE

Claude Code is introducing a new "Tool Search" feature for its MCP protocol to address the challenge of managing a growing number of tools. Previously, numerous MCP tool descriptions could consume a lot of context. This new feature dynamically loads tools via search only when needed.

AGENT SKILLS

Agent Skills is Vercel's collection of packaged instructions and scripts for AI coding agents. For example, the `react-best-practices` skill provides over 40 rules for optimizing React and Next.js performance across various categories like bundle size and data fetching.

JSON RENDER

JSON Render is a library that gives AI the ability to safely generate UI, such as dashboards and widgets, from prompts. It requires devs to define a catalog of allowed components and actions, which constrains the AI's output to predictable, schema-validated JSON. This guardrailed approach makes sure the AI-generated UI is always predictable, secure, and can stream and render progressively.

GETTING REAL LEVERAGE FROM CLAUDE CODE

Traditional Git worktrees alone when using Claude Code are not enough as they lead to shared resource conflicts like port contention and database clashes. Instead, it's better to use a local Kubernetes cluster (k3d) to provide true namespace isolation for each worktree, with Claude assisting in writing the necessary, often tedious, YAML configurations.

THE QUESTION YOUR OBSERVABILITY VENDOR WON'T ANSWER

In observability, vendors strong-arm clients and costs constantly escalate, forcing users to act as cost police. The core issue is vendors' willful ignorance of how much observability data is waste. An average of 40% of observability data is not used.

THE UNBEARABLE FRUSTRATION OF FIGURING OUT APIS

LLMS ARE A 400-YEAR-LONG CONFIDENCE TRICK

LLMs are a trillion-dollar confidence trick, using humanity's centuries-long trust in machine accuracy and fear of being left behind, to push adoption.

JUNIOR DEVELOPERS IN THE AGE OF AI

Despite companies blaming AI for the mass dismissal of junior engineers, these hires are actually necessary for building institutional knowledge and even accelerating AI adoption.

DEBUGGING NEXT.JS WITHOUT ERRORS: HOW LOGS REVEALED A PRODUCTION BUG

Logs are necessary for debugging application problems that don't throw explicit errors.

CLAUDE COWORK EXFILTRATES FILES

Claude Cowork is vulnerable to file exfiltration via indirect prompt injection that exploits known-but-unresolved isolation flaws, allowing user files to be uploaded directly to an attacker's Anthropic account.

Try it free for 14 days →

WHY WE BUILT OUR OWN BACKGROUND AGENT

Inspect is a custom background coding agent developed by Ramp that autonomously writes, verifies, and debugs code with the same context and tools a human engineer would use. The agent operates in fast, sandboxed VMs with full development environments, integrates with many existing tools, and supports interfaces like Slack, web, and Chrome extensions. Inspect was adopted quickly internally, writing approximately 30% of merged pull requests.

HOW A 40-LINE FIX ELIMINATED A 400X PERFORMANCE GAP

A method in OpenJDK on Linux previously suffered from a severe 30x-400x performance gap. The slower implementation was due to standards requiring a less direct approach for user CPU time, unlike the faster `clock_gettime()` function used for total CPU time. A recent 40-line fix exploited an obscure yet stable Linux kernel detail to enable `clock_gettime()` to directly retrieve user-time-only, improving performance a lot.

REACT HAS FINALLY SOLVED ITS BIGGEST PROBLEM: THE JOYS OF USEEFFECTEVENT

React has finally introduced the `useEffectEvent` Hook to solve its biggest problem: infinite loops and stale closures caused by `useEffect` dependency arrays. The hook provides a cleaner solution than the traditional `useRef` workaround by allowing devs to access current state values within effects without triggering re-runs when those values change.

NO MANAGEMENT NEEDED: ANTI-PATTERNS IN EARLY-STAGE ENGINEERING TEAMS

For early-stage founders facing engineering management issues, the optimal solution is often to do nothing, focusing instead on building the product and engaging with users. Attempting to actively "motivate" engineers (who should be inherently motivated hires), hiring dedicated managers prematurely, and copying complex management practices from large companies like Google are the wrong things to do.

HOW THE HELL ARE YOU SUPPOSED TO HAVE A CAREER IN TECH IN 2026?

The tech industry has had half a million workers laid off since ChatGPT's release, despite billions in AI investments. Tech workers need to understand their role within organizational systems rather than just focusing on technical skills, recognizing that companies increasingly view code and specific expertise as replaceable means to an end.

DICER

Dicer is Databricks' auto-sharder, a foundational infrastructure system for building sharded services with low latency, high availability, and cost efficiency at scale. It addresses the inefficiencies of stateless models by colocating in-memory state with computation, so there are no issues like network latency and unnecessary data fetching.

SWARK

Swark is a free and open-source VS Code extension that automatically generates architecture diagrams from code using LLMs. It integrates directly with GitHub Copilot and outputs diagrams in the Mermaid.js format.

OPENCODE VS CLAUDE CODE

OpenCode and Claude Code are both AI coding assistants that let you chat with your codebase and run terminal commands. Claude Code offers a polished, integrated experience locked to Anthropic's ecosystem, while OpenCode is an open-source alternative supporting 75+ AI providers. In head-to-head testing using the same model, Claude Code was faster and more streamlined (9 minutes vs 16 minutes total), while OpenCode was more thorough.

EVERY GITHUB OBJECT HAS TWO IDS

When building a feature for Greptile, this dev encountered an issue generating clickable GitHub links because their system stored GraphQL `node IDs` (e.g., `PRRC_kw...`), while web URLs required integer database IDs. Further investigation showed that GitHub uses two ID formats: an older, string-based system for legacy objects and a newer, more complex system for modern objects, despite inconsistent application.

NPM TO IMPLEMENT STAGED PUBLISHING AFTER TURBULENT SHIFT OFF...

npm is implementing staged publishing, a new feature that will introduce a review window before packages go live. This will require MFA-verified approval from package owners to catch malicious changes before they spread. This comes after the 2025 Shai-Hulud supply chain attacks and a disruptive migration away from classic npm tokens in December that caused workflow disruptions for maintainers.

of actually deploying it there. Learn more about mirrord

NOGIC

Nogic is a Visual Studio Code extension that helps users visualize their codebase structure with interactive diagrams.

MAKING AI GREAT AT FRONTEND

Some clever AI prompting techniques can make the frontend output of AI much better, such as sharing progressive references and using plan mode.

USING PROXIES TO HIDE SECRETS FROM CLAUDE CODE

Claude Code has security challenges, as its sandboxes can expose sensitive data such as API keys through network access or environment variables, but network proxies can be used to inject the actual API keys in a safer manner.

WORK EXPANDS. TIME VANISHES. HERE'S WHY

Parkinson's and Hofstadter's Laws say that work expands to fill available time and always takes longer than expected, but this can be somewhat mitigated with timeboxing and clearer definitions of the work.

Get a free developer account →

HOW TO STEAL ANY REACT COMPONENT

It's possible to "steal" or reconstruct any React component from a production website, even without access to its source code. React Fiber is an internal tree structure in the browser that exposes which components create specific DOM elements and what props they receive. You can generate a component's code by collecting multiple examples of a component's props and its resulting HTML and feeding it into an LLM.

HOW TO BUILD AGENTS WITH FILESYSTEMS AND BASH

A simple, effective architecture for AI agents is a standard filesystem and Bash tools. Since LLMs have a native understanding of code and Unix commands, agents can easily navigate and retrieve information from data structured as files, much like exploring a codebase. This approach overcomes the limitations of prompt stuffing and imprecise vector search by offering natural data hierarchies, exact retrieval, and minimal context loading.

REPLACING LLM DATE MATH WITH DETERMINISTIC SCRIPTS TO IMPROVE CALENDAR ACCURACY

LLMs like Claude frequently fail at date math, causing calendar errors in tools like gcalcli. Offloading calculations to deterministic scripts ensures accuracy. Developers can eliminate hallucinations and create reliable, repeatable AI workflows by updating CLAUDE.md instructions to execute code instead of manual reasoning.

ANTHROPIC MADE A BIG MISTAKE

Anthropic recently made a business blunder by closing a loophole that allowed customers to use their cheaper subscription models with popular third-party coding agents like OpenCode. This move alienated many paying customers who felt they were unjustly restricted from using a service they already purchased. OpenAI capitalized on this by officially supporting its own subscriptions within these same third-party agents.

THE GREAT FILTER (OR WHY HIGH PERFORMANCE STILL ELUDES MOST DEV TEAMS, EVEN WITH AI)

Despite the rise of AI-assisted coding, there's no widespread evidence of increased developer productivity or business value. DORA data shows that only already high-performing teams see benefits, while AI often slows down inefficient ones. The key differentiator for high-performing teams is addressing development process bottlenecks by adopting continuous, small-batch software work with tight feedback loops.

INTRODUCING COWORK

Claude Cowork is a new feature that extends Claude Code to allow anyone, not just developers, to interact with Claude by giving it direct access to files on their computer. This allows Claude to read, edit, or create files within a designated folder, completing tasks like organizing documents or drafting reports with enhanced agency.

AGENT OF EMPIRES

Agent of Empires is a Rust-written terminal session manager for Linux and macOS. It functions as a wrapper around tmux, creating a separate tmux session for each AI coding project and ensuring sessions persist reliably. The tool has a TUI dashboard for visual session management, including creation, attachment, detachment, deletion, and group organization.