Google’s decision model scores 1,000 questions at once

Welcome back. Ever since Jev, big tech has been obsessed with decision models. Google just dropped its best one yet, capable of evaluating 1,000 questions in a single run. Meanwhile, a new phone built for AI agents has devs asking whether the iPhone era is ending. Bold claim, but it’s certainly got our attention.

Also: A site tracking tech layoffs and job openings live, a 4-step Claude Code setup that gets the best out of each model, and why the waterfall approach of building deserves a comeback.

Google’s decision model scores 1,000 questions at once: The search giant just dropped Gemini Decision 1, a multimodal model built for closed-set decisions. Feed it text, images, audio, video or PDFs, and it scores up to 1,000 questions in one pass without generating any output tokens. Your code acts directly on the probabilities it returns. Typical responses take 150–250ms, and it costs $0.10 per million input tokens with no output charges. Check out how you can start building here or read the API docs.

This AI phone will have more context about you than your iPhone: The Y Combinator-backed startup just launched the public beta of Blue Phone, built on Google’s Pixel 10 Pro and Pro XL. It runs Blue OS, a custom Android system that lets AI agents handle multiple tasks across your apps at once. Instead of users tapping through your UI, an agent could soon be doing it for you. For devs, it’s another reason to start making apps agent-friendly. Priced at $1,299, you can join the launch here.

AWS gives coding agents their own local cloud: The cloud giant just shipped AWS Emulator, a local stand-in for over 200 of its services. Coding agents write code faster than teams can test it, and checking every change against the real cloud adds delays and costs. Now you and your agents can run that test loop offline without paying for the cloud. Fittingly, autonomous agents wrote the entire emulator. But keep in mind AWS’s own docs say it can’t yet simulate throttling, outages, or regional differences. Try it here.

CodeAF's senior-dev specialist finished #1 of ten coding harnesses on DeepSWE on one open model, at the lowest cost per solved issue.

Most coding agents are still one linear chat with subagents bolted on, so you end up as the manager. CodeAF replaces that chat with an org:

Open source, built for open models.

Try CodeAF on GitHub

The old rules are out. Throughout the 2010s, moving into engineering management meant leaving the code behind. You traded technical decisions for meetings, hiring and keeping teams on track. But AI is changing how software gets built, and managers who’ve spent years away from the code are finding themselves at a disadvantage.

A warning from the top. Will Larson, CTO of Imprint and former CTO of Carta and Calm, told The Pragmatic Engineer he hasn’t seen a shift this big since he started in 2008. He warns that middle management as we know it might not be valued three years from now. The choice is to either reskill or retire.

He chose to reskill. Larson swapped a roughly 250-person org at Carta for a 25-person team at Imprint. Since then, he’s merged over hundreds of PRs into production, more than the rest of his career combined. Here’s how you can follow his lead:

Plot twist. Larson says managers who feel betrayed have a point. They spent years climbing the ladder, only for the rules to change underneath them. But he also thinks the new way of working is a lot more fun. Catch the full conversation here.

A TV, a coworker, someone in the car: your STT transcribes them all, and your agent answers the wrong person. Krisp Voice Isolation keeps only the caller.

• 73% fewer word errors across 265 real recordings

• Works in front of any STT, 15 ms on CPU

• API and SDK, already in Vapi as Smart Denoising

Get $30 in free API credits →

P.S. You are the reason our team spends hundreds of hours each week tracking down the best stories for you. Please take a moment to reply to this email with your favorite story from this section so we can cover more of what helps you most in your career. 

How to run Claude Code, Codex, and Grok from one app: This tutorial shows how to use T3 Code as a control room for all your coding agents. You’ll learn how to switch between models and accounts, hand context across threads, use worktrees and skills, and manage multiple agent workflows without living in five different terminals.

Ace: GitHub Next's open-source tool lets teams and agents collaborate on the same Mac. Jump into shared channels, steer each other's agents, and give them direct access to your local shell, tools, and apps over Tailscale rather than a hosted sandbox. 

Prompting guide for Claude Fable 5.2 (by Anthropic): This guide breaks down the prompting tweaks that matter most for the new model. You’ll learn how to tune effort, stop long tasks from ending early, keep approvals and subagents under control, and update older prompts or harness rules that can now make Fable slower, costlier, or more cautious than it needs to be.

An implicit state machine is state logic that is scattered across your code instead of written out in one place. The result is your app can enter states that should be impossible, and those bugs ship even when tests pass. XState creator David Khourshid has a full explanation on what they are and why you need to be wary of them.

Steve Ruiz, creator of tldraw, shared a one-line prompt that finds them.

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 400K+ 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