Google gives devs a faster coding model

Welcome back. Google’s on a roll, and its latest model is built to make your everyday coding tasks faster and cheaper. Meanwhile, Anthropic is giving you more control over Claude’s agents, letting you run the entire stack, from sandboxes to session history, in your own cloud or on-premises.

Also: Check if your coding agents hog resources, a researcher’s “proof” that four Chinese AI labs used banned NVIDIA chips, and seven steps to turn AI slop into gold.

Google gives devs a faster coding model: The search giant just dropped Gemini 4 Hydrogen, a lighter model built for everyday coding tasks you hand off to agents. Google says it finishes jobs 2.5x quicker than Gemini 3.8 Flash and scores three times higher on Terminal-Bench 4.1. And when you need more muscle, a new Orchestrator in Antigravity pairs it with the heavier Carbon model, delivering near-Carbon results on 24% less quota. Pricing is a flat $1.50/$7.50 per 1M input/output tokens. Try it here.

Anthropic lets you run its agents behind your firewall: The AI lab just opened a private beta of self-hosted Claude Managed Agents, so your team can run the entire agent stack in its own cloud or on-premise. That includes sandboxes, credentials and session history, making it easier to put agents to work fixing builds and triaging alerts on sensitive code. Anthropic is also rolling out Claude Meetings in beta, which joins your video calls, creates docs and whiteboards as you talk, and saves notes afterward.

Liquid AI makes on-device voice agents faster: The AI startup just shipped LiquidCodec, a ~74M-parameter model that turns speech into tokens language models can work with, then converts them back into audio. Liquid AI says it encodes speech 361x faster than real time on a single CPU thread, making it roughly 13x faster than Kyutai’s Mimi. That could mean quicker responses and smoother streaming for your voice agents, even on edge devices. It runs through Hugging Face or llama.cpp. Try it here.

AI agents now write more of your code than ever, which makes reviewing it the new bottleneck.

Air, now built into JetBrains IDEs, gives you one place to run your agents and stay in control of every change:

Try Air with free AI

Five VMs, one problem. Gumloop’s AI agents had gotten slow to start, and users noticed. CTO Rahul Behal then opened five VMs and asked a coding agent in each one to find the cause. Before he finished setting up the fifth, the first came back with a fix. Most of the time, the agents were right. So why was he still in the middle?

Enter the factory. That experiment became the basis of Gumloop’s Software Factory, where agents handle the repetitive work and engineers focus on what actually needs their attention. Here’s the playbook in six steps:

The scoreboard. In six weeks, engineers merged 571 agent-written PRs. P99 agent start times dropped from 24.9 to 4.5 seconds, and 83% of closed agent-written PRs needed zero human code changes.

Write it down. Gumloop wrote or updated 94 skills in five weeks, and no-human-edit merges jumped from 57% to 89%. The more you document your team’s decisions, the less you have to babysit agents. The full breakdown goes deeper, from Slack judgment calls to adding Devin after a Cursor outage.

Granola Voice Typing is now available on Mac, Windows and iOS (Android coming soon).

Hold a key, talk, and clean text lands in any app, with your colleagues' names and jargon spelled right because Granola already knows your meetings. Unlimited words for all Granola users.

First month on us with code THECODE.

How to manage coding agents across multiple repos: Senior dev Matt Pocock shows how to use a single “Home Repo” to keep shared skills, steering files, and custom tooling in one place instead of duplicating them across every project. You’ll learn how to structure the setup, scale it across repos, and keep your agent workflow consistent without moving everything into a monorepo.

Firecrawl: Turn ChatGPT or Claude into a job-hunting agent. It can match your resume against 7M+ open roles, find hiring managers across 250M+ people, and surface what they care about before you reach out. It also works with Claude Code, Codex, Cursor, Grok Build, OpenCode, and other agent setups.

How to train a transformer with no ML experience: This guide follows a software engineer building and training a transformer from scratch for a real coding challenge. You’ll learn how to turn application state into tokens, use distillation and self-play, shrink the model to fit tight runtime limits, and test whether better training scores actually translate into better results in the real system.

A design engineer shared a QA prompt that fixes a common failure: AI-built frontends look fine on your desktop and fall apart on a phone with bad signal.

Claude opens the app, resizes the viewport, throttles the network, patches what breaks, and hands back before/after screenshots as receipts. The same MCP works in Cursor and Gemini CLI.

A software engineer shared a site tracking tech layoffs and 90K+ open roles in real time, and it’s already blown up.

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