Welcome back. AI has split engineering into two distinct camps. On one side, developers use LLMs to automate their entire workflow, barely touching their editors. On the other, true craftsmen burn out managing the sloppy code being shipped. One VC put the phenomenon into words, and it's worth the read.

Also: How a bootstrapped founder ships code from any device, which of five archetypes the Claude Code creator says you'll become, and four ways an Intuit engineer breaks into tech.

OpenAI previews its strongest models but limits access: The ChatGPT maker just previewed GPT-5.6, a three-tier lineup led by Sol, its most advanced model for coding and cybersecurity. But a government request has locked all three behind a private preview for a few vetted partners on the API and Codex. The lab is pushing back, warning the gatekeeping puts essential tools out of developers' hands. A wider release is expected within weeks.

Ex-Databricks AI chief targets 1,000x cheaper inference: A stealth startup founded by former Databricks AI chief Naveen Rao just dropped Un-0. It’s an image generator that reportedly matches leading diffusion models. The catch is that it runs on a completely different oscillator-based architecture rather than GPUs. Rao claims real silicon could eventually cut power draw a thousandfold. For teams drowning in compute bills, it's one to watch.

Chinese models are catching up to America's frontier labs: Two recent releases just shook up the AI race. Qihoo 360's Tulongfeng, which its founder calls a "Chinese version of Mythos," reportedly matches Anthropic at spotting cyber bugs. Meanwhile, Z.ai's open-weight GLM-5.2 rivals top US models at a fraction of the cost and is already among OpenRouter's most-used. For engineering teams, capable models are getting cheaper. For security pros, that same power is spreading fast.

Your customer lives in the real world, and it’s messy: intermittent connections, mid-onboarding drop-offs, edge cases on devices you've never tested. Most observability tools never see any of it. bitdrift captures 100% of mobile data, unsampled and in real time, so it’s immediately queryable by engineers and AI agents.

Query reality with bitdrift: mobile observability built for the real world.

The agent backlash. Open up almost any dev tool these days and you'll find an AI agent waiting to do the work for you. But last weekend, YC-backed founder Rhys Sullivan argued that engineers are less likely to use these agents. He’d prefer to plug specialized skills and APIs into his existing daily agent rather than deal with a new one. If you've just launched a standalone agent, your power users are likely already tuning it out.

The losing race. The gut reaction is to try and win them back by improving your agents. You might be drawn towards a more powerful model, local file access, better polish, etc. But you just can't win that race. The best models are way too expensive to give away for free to every user, so built-in agents inevitably drop down to cheaper, weaker ones. Even then, they can't actually touch the files and repos on a developer's machine. You're basically paying to build a watered-down version of what they already have. 

Open the product up. Stop competing with the agent they trust, and feed it instead. Give it docs it can parse, a way to call your API, and an MCP server, the open standard any agent plugs into without custom wiring. Another thing: the same server runs your own in-app agent too. Build the primitives once, point your agent at them, and let power users point theirs at the same place, so you maintain one system instead of two.

Cloudflare is doing exactly that. They bundled their entire API into one MCP server any agent can call, and it fits in about 1,000 tokens of context against the million-plus it would take to expose every endpoint as its own tool.

Where to start. There’s no need to start from scratch. Sullivan suggests shipping an MCP server with your product guides and letting users know they can stick with their preferred agent. Be sure to lock it down with authentication to keep your data secure. To stay competitive, focus on your proprietary data and a usage-based pricing model. Without those, you're just building a utility and risking becoming a commodity.

SmartBear Reflect is an AI-powered, codeless test automation platform that turns plain-English test steps into automated actions—no coding required. 

Try it at no cost.

Reviewers often get stuck on a big diff when they don't understand the logic behind it. This usually leads to long delays or a lazy rubber-stamp approval.

Claude Code's artifacts feature fixes this by building a walkthrough directly from your diff and the surrounding code, then gives you a shareable link. Just run this inside your session: 

Claude looks at your actual import graph and diff instead of just a summary, so the walkthrough accurately reflects the real changes. You can share it right from the page header.

Every artifact stays private until you decide to share it, and it's only viewable by authenticated members of your organization.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

When to use Claude Agent SDK over Claude Code: You’ll learn how the Claude Agent SDK compares to the Claude Code CLI. We will build a multi-agent language tutor in Python. You’ll also find out how to manage complex AI setups in parallel and create production-ready apps with built-in testing tools. 

Ports (by a Cursor engineer): This macOS menu bar app tracks local dev servers, showing ports, resource usage, and uptime. It lets you quickly kill zombie processes and open terminals to keep your workflow clean.

Strix (27.1K ⭐): AI agents that find and validate vulnerabilities with real proof-of-concepts. It offers fast, accurate security testing for developers without the fluff of manual pentesting. 

Using local coding agents: Paid coding assistants like Claude Code and Codex come with subscription fees and data privacy risks. This guide shows that running open-weight models locally with Ollama delivers similar performance for free while keeping your data completely private. 

A full-stack engineer used Claude to build a token-based design system that scales automatically. 

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team