DeepSeek lets you build agent plugins with a prompt

Welcome back. It turns out OpenAI’s agents were pushing so much code that the company ended up rebuilding its own Git backend. Now, they’re sharing what they learned along the way. Meanwhile, DeepSeek wants to make building plugins as easy as typing a prompt.

Also: MIT researchers make a tiny but surprising edit to an AI model’s weights, an AI phone aiming to replace your iPhone, and a 12-stage AI security engineering roadmap.

DeepSeek lets you build agent plugins with a prompt: The Chinese AI lab just rolled out Creator Mode for DeepSeek Harness, letting you create plugins simply by describing what you want in a text prompt. While models decide what to do, a harness is what gets the job done. DeepSeek’s open-source Harness takes an “everything is a plugin” approach, so whether you need a tool, skill, or interface tweak, the agent will build it for you. It’s still in preview, though, so expect API changes. Try it here.

OpenAI rebuilds Git to keep up with its agents: The ChatGPT maker’s AI agents were pushing so much code that its Git setup started to buckle, with reliability sometimes dipping below 90%. So OpenAI built Pando, a new server designed to handle all that parallel agent traffic without changing how engineers or agents use Git. And it’s paying off. The company says pushing code now takes nearly 80% less time. Pando is staying internal for now, but OpenAI is sharing its Git fixes with the community.

Hume wants to replace your audio model stack: The voice AI startup just unveiled its Audio Decision API model, letting you ask questions about a recording in plain language. That means figuring out how many people are speaking, whether a caller sounds frustrated, or even if someone laughed. Instead of stitching together 10+ models, you now need one API call on live audio or hours of recordings. Trained on 10M+ human judgments, it scored 82% on the company’s benchmark, edging out Gemini 3.8 Flash at 78%. Join the waitlist.

Claude Code, Cursor, Codex. The output ceiling is insane. But "fix the bug" gets you slop. Detailed context gets you shipping code.

The problem: typing out file references, edge cases, and architecture context 40 times a day is exhausting. So you cut corners. Your agent gets lazy input.

Wispr Flow lets you speak full prompts with syntax-aware dictation. camelCase, snake_case, file names all preserved. Auto-tags files in Cursor and Windsurf.

Used by engineers at OpenAI, Vercel, and Clay. Works on Mac, Windows, iPhone, and Android.

Better prompts in. Better code out. Try Free

Speed at what cost. Proton, the Swiss company behind Proton Mail, is going all in on AI and things are getting messy. A 404 Media investigation published last week revealed that CEO Andy Yen dismissed an engineer’s request to pentest new code as “loser thinking.” And that was just the beginning. He also told staff that 2026 would be the last year Proton hires engineers and announced AI re-interviews that could decide who keeps their job. Unsurprisingly, that didn’t go down well.

A quick U-turn. Users started canceling subscriptions and calling for Yen’s resignation. A day later, he scrapped the re-interviews, admitted a communication failure on his part, and Proton apologized publicly. But the AI push isn’t going anywhere. Employees say the re-interview criteria are being folded into performance reviews.

The real takeaway. CTO Bart Butler’s leaked manifesto is worth your attention. Here are the ideas you can put to work:

Two sides to it. Reactions have been mixed. Ruby on Rails creator DHH defended the push, while Clean Code author Robert “Uncle Bob” Martin echoed the manifesto’s core argument. Critics, meanwhile, focused on Proton’s workplace culture. Either way, if you’re an engineer in 2026, there’s plenty worth borrowing from Butler’s full manifesto.

How to build your own Claude Code software factory: This tutorial shows how to stop babysitting coding agents and build a workflow where a linear ticket kicks off the work, multiple agents build in parallel, and you only step in when a decision actually needs you. You’ll learn how to combine Claude Code, T3 Code, PStack, cloud sandboxes, playbooks, and verification into one repeatable setup.

NanoUI: Build tiny cross-platform desktop apps in Zig with either native GPU-drawn UI or a web frontend. It runs on macOS, Windows, and Linux, ships as a single binary, and includes headless testing and screenshots so coding agents can build and verify interfaces without opening a window.

How to build a design system for coding agents: This guide shows how to turn the design decisions in your head into a system an agent can reliably use. You’ll learn how to structure primitives, semantic tokens, components, reusable blocks, rules, and skills so agents stop guessing and start building interfaces that actually feel like your product.

Tell Claude Code to “make it faster” and it usually adds a cache. A dev shared a perf-review prompt that names the exact structures to ask for instead.

Expect a report with files, lines, and estimated speedups. Approve the changes you want, skip the rest.

Anthropic is expanding watermarks across Claude outputs starting Nov. 3, and this post breaks down what you need to watch for.

If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:

We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.

Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ 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