TLDR Dev 2026-02-20

You can't fix QA by throwing more bodies at it – not when everyone is coding 10× faster with AI.

You need an AI counterweight that validates every release on time.

Teams at Nordea, Pricer, and Upsales use QA.tech to cut thousands of hours of manual testing each month, reduce QA spend, and ship faster.

With QA.tech, you get:

> AI testing agents that scale with your development.

> Shift-left testing that keeps release control in-house.

> No QA bottlenecks.

🎁 Special TLDR offer:The first 20 teams to start a PoC receive up to $20,000 in free credits. Get started at https://qa.tech/try-now/tldr

Four AI coding tools had their context windows and token usage analyzed while fixing a specific bug in Express.js. Despite all tools successfully solving the bug, their token consumption varied a lot, with Gemini Pro using up to 350,000 tokens compared to Claude Opus's 23,000. Each tool used a distinct strategy: Opus was surgical but carried high tool definition overhead, Sonnet was methodical, Codex efficiently used Unix-like commands, and Gemini aggressively dumped vast amounts of data into its context.

Modern AI agent frameworks in Python and TypeScript are similar to the Elixir/Erlang BEAM virtual machine, which solved similar concurrency and fault-recovery challenges for telecom in 1986. The BEAM's actor model has features like isolated lightweight processes, message passing, and supervision trees, well-suited for the long-lived, non-deterministic interactions of AI agents.

This DevRel engineer is ending a 7.5-year career at Google after being forcibly re-organized onto a new team focused on AI products, which conflicted with her moral values and long-standing dedication to Firebase. This new role also involved a manager with whom she felt psychologically unsafe and untrusting. She took severance to leave.

AI should be conceptualized as an exoskeleton that amplifies human capabilities, rather than an autonomous coworker. Physical exoskeletons enhance strength, reduce injuries, and improve endurance in fields like manufacturing and military, without replacing the human operator. AI should be used the same way.

Expo's web team committed 100% to Claude Code for a month and ended up with wins, losses, and lessons you can read about in this blog. They also added the ability to prompt Builds and Workflows with Claude with Expo MCP Server. Try the new MCP server

cmux is a native macOS terminal application built with Swift and AppKit that extends the Ghostty experience with vertical tabs and GPU acceleration. It has an intelligent notification system that visually alerts users when AI coding agents require attention, along with a sidebar displaying vital context like git branches and listening ports.

Google has released Gemini 3.1 Pro, an upgraded core intelligence model designed for complex problem-solving and advanced reasoning. This new version has more than double the reasoning performance of Gemini 3 Pro, and it is specifically built for tasks where simple answers are insufficient.

Step 3.5 Flash is a frontier open-source foundation model by StepFun that offers efficient, high-speed reasoning and agentic capabilities with tool use, long context, and local deployment.

Anthropic's research analyzed millions of human-AI agent interactions across Claude Code and its public API to understand how agents are used in real-world contexts. It found that agents like Claude Code operate autonomously for longer periods, and experienced users grant them more independence (auto-approving more). Claude Code also proactively pauses for clarification more often than humans interrupt it on complex tasks.

AI systems like Aristotle solve complex mathematical proofs, while tools like OpenClaw reduce software development tasks from days to minutes. This exponential feedback loop is similar to the Industrial Revolution, requiring a fundamental rethink of economic structures to manage rapid automation.

AI accelerates software divergence, making change cheaper than commitment.

Relying on AI for creative and intellectual work bypasses the deep engagement and prolonged thought processes necessary for originality, ultimately leading to shallow ideas and uninteresting projects.

fff.nvim is a Neovim plugin that is a fast file finder, using a dedicated Rust backend for typo-resistant fuzzy search, live grep, Git status integration, and image previews.

A solo founder recounts his decision to migrate his web application from Rust to Node.js due to Rust's current limitations in web-specific features, ecosystem maturity, and compilation speed.