Welcome back. If you've been leaning hard on coding agents lately, you've probably hit that context window wall. You're definitely not the only one. The lead engineer at Codex just dropped his personal setup for stretching context in heavy coding sessions, and devs are already swearing by it. Details below. 

Also: How a VP of Engineering promotes engineers, Andrew Ng explains what AI engineers actually need, and how top devs use GrokBot.

Alibaba’s open models are leaving Meta and Google far behind: The Chinese tech giant just released the weights for a compact 27B model that is small enough to run on your MacBook. According to coding benchmarks, the lab claims it performs on par with Anthropic's Opus 4.6 Max. The Qwen family just topped 3 billion downloads in only six months, making it the most popular open model family, far surpassing Google's 418 million and Meta's 227 million. You can grab the weights here.

Anthropic watermarks Claude text but spares most code: The AI lab just dropped a blog post detailing how their new watermarks will work. And they’re promising there’s no catch for devs: no extra tokens, no hidden costs, and zero performance lag. Your code should stay virtually untouched too, since the watermarking only affects word choice, and strict syntax doesn't leave much room for changes. The move comes just as Anthropic CEO defended his recent AI warnings in a rare social media post. Read his statement here.

OpenAI engineer reveals 1M-token setup for Codex: A post by an OpenAI researcher just went viral after he shared how to enable a 1M-token context window in Codex for GPT-5.6 Sol. With just a few lines in a config file or a single command flag, developers can now massively expand the model's immediate memory. This lets the agent track more code, tool outputs, and conversation history before it ever has to start cutting older data. Check out the setup here.

AI is writing more and more code than ever before in history. 

So why are you still reviewing PRs like it’s 2010—alphabetical order, zero context, no high-level summary of why changes were made?

CodeRabbit’s Change Stack restructures PR reviews into layered walkthroughs, giving you…

And if you need help? Just ask the Agent Chat for context directly where you’re working.  

Try it here (early access costs nothing).

The big debate. When in doubt, hire a Forward Deployed Engineer (FDE). Or at least, that’s the trend Decagon CEO Jesse Zhang is noticing across the AI landscape. Demand for the role is surging, with labs like Anthropic and OpenAI building out dedicated deployment teams. But an FDE is a major investment, so the real question according to Zhang becomes whether your company actually needs one.

The field work. The defense of the role comes down to discovery. When a category is brand new, there's no playbook, so engineers work out how the product should behave alongside the customer. This is where an FDE earns its place:

But this model can easily slip into a services business. Here are a few red flags that you’ve crossed that line:

Watch the numbers. Every deployment should make the next one cheaper. FDEs earn their place while they're still uncovering the workflow, and they turn into overhead once they're only absorbing it. If each rollout keeps costing the same, discovery is over and you're drifting into a services business.

PS: Here’s Zhang's full essay and also another view from Andrew Ng who further pushes back, arguing the field will need far more AI engineers than FDEs.

How to build your own AgentOS on Claude's Agent SDK: You will learn how to build practical, LLM-powered applications while mastering prompt techniques, model chaining, and the creation of smart AI agents. This tutorial equips you to rapidly prototype, automate complex workflows, and level up your software projects. 

DeepSeek Harness (114.2K ⭐): An open-source agent harness from DeepSeek AI, built on Cordis. Its big idea is simple: everything is a plugin, so models, tools, workflows, and interfaces can all be swapped out easily.

How to run Claude Code without burning tokens (by Anthropic): Coding costs used to be fixed, regardless of the task. Today, the price for the exact same job can fluctuate wildly depending on your workflow. This guide helps you keep your expenses under control.

You hit your limit at 11 PM right in the middle of a refactor, and the session just stops. Even though the window resets at 3 AM, nobody’s awake to hit "continue," so you end up losing four hours of progress on work that was ready to go. Anthropic just shipped a fix for this: 

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

Check out how this viral workflow cuts token costs for Claude and Codex by 95% by offloading the grunt work.

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