Gemini 3.8 Flash goes all-in on agents

Welcome back. Google is on a “flash streak”. Just weeks after the launch of its fastest model, the search giant has now released two new updates. Meanwhile, an AI startup revealed how Claude Code and Codex costs 30x more for the same coding tasks than other alternatives. Read details below.

Also: How Claude Code fixed a broken TV, how Uber used an employee’s anonymous feedback against them for firing, and a prompt to delete over-engineered code.

Google unveils two models on the same day: The search giant dropped Gemini 3.8 Flash, its smartest model yet. It’s built to outwork larger models across coding, agentic workflows, and multi-step reasoning. Whether you prefer speed or savings, it doesn’t matter. You can dial the effort down or stay on 3.7 Flash for efficiency-first work. The initial reports on the pricing suggest just $0.75 per million input tokens. Google also announced Gemini 3.8 Flash Cyber, built to hunt down security bugs. The company says it produces 2.6 times more correct Chrome patches than larger models. Try it here.

Pi coding harness won't break the bank like Codex or Claude Code: AI startup Runta just dropped a new benchmark, testing nine coding harnesses on the same model to see who's actually delivering. While pass rates were pretty close across the board, the cost gap was massive. Some harnesses cost 17.5x more than others. Pi harness crushed it by fixing a tough bug for just $2.50, while Claude Code burned through $64.36 for the same task. It turns out your harness choice matters just as much as your model choice if you want to ship without breaking the bank.

Meta takes its biggest model leap yet: The social media giant just dropped Muse Spark 1.3 through Muse Code and the Meta Model API. Chief AI Officer Alexandr Wang is calling it their biggest performance jump yet, especially for coding and agentic workflows. It’s 25% more token-efficient and handles complex details way better than older versions. Plus, it’s smart enough to know its limits and double-checks with you before doing anything permanent. Start building.

Trigger.dev runs your AI agents as durable tasks inside your codebase, in plain TypeScript with full control of the runtime. No serverless timeouts, so even long agents run to completion, surviving refreshes, redeploys, and crashes along the way.

Write your logic, test it, and deploy agents that scale on demand.

Start building agents here.

The ceiling nobody talks about. Off-the-shelf coding agents like Claude Code or Codex are great on day one. But they often stall inside real companies because they run locally, which usually limits them to fewer tasks at a time and cuts them off from internal systems. A deep-dive by veteran engineer Gergely Orosz shows how Ramp got past this.

Do it yourself. Ramp built Inspect, a background agent that now handles about 75% of its merged pull requests. Each session runs in a remote sandbox set up like a dev machine, with the same internal access as a Ramp engineer.

Where it shows up. That access is what makes Inspect useful across the whole workflow:

The takeaway. The agent alone isn't the breakthrough. It uses OpenCode, which anyone can install. The real secret is the sandbox: every Inspect session boots a nearly identical copy of a Ramp dev machine, giving it internal access most tools never get. That's the true value of building your own agent. It bridges the gap from the first idea to finished code. Ramp's full engineering write-up shows how they pulled it off.

What is the best agent harness and how to use them: In this tutorial developers can explore top agent harnesses like Pi, OMP, and Zcode. You’ll learn how to integrate these frameworks into your workflows, automate non-coding tasks, build custom agent setups, and boost daily productivity using AI coding environments.

Boop: A self-hosted alternative to Sentry that sends app error and event alerts straight to your phone as native iOS notifications, using 8 MB of memory.

5 things AI engineer should know about agent sandboxes (by Google): This cookbook breaks down why cold-start benchmarks mislead and why network egress, not the hypervisor, is the real attack surface for agents. Includes a 4-question rubric for picking a sandbox stack.

Your Codex and Claude quotas sit idle while you work in one tool. This dev’s fix: install the agents inside Grok Bot's cloud machine and delegate from one chat.

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

A designer just dropped a library that adds a human touch back to polished UIs. One quick install, and your interface feels way less AI-made.

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