Welcome back. Frontier AI models are turning into guarded national assets. Months ago, Anthropic blocked Chinese firms from using Claude. Now, Reuters reported that Beijing is considering a similar move. They may soon restrict foreign access to top models from Alibaba, ByteDance and Z.ai.

Also: A senior engineer's fix for context bloating, how to become an Applied AI Engineer, and how to manage engineering teams in the AI era.

Anthropic brings Claude Cowork to your phone and extends Fable 5 access: The AI lab just rolled out Cowork beta on web and mobile for Max subscribers. You can kick off a task at your desk, track it from your phone, and access the final output from any device. The update also lets you set “scheduled tasks” that will run even if your computer is closed. Plus, Fable 5 stays in all paid plans through July 12 before moving to usage credits.

Google preps Gemini's managed agents for production work: The search giant just dropped four upgrades for its hosted agents in the Gemini API. MCP support lets them talk directly to internal databases and APIs, cutting out the need for custom proxy middleware. Background execution keeps long-running tasks on the server while your app checks in for results. Custom function calling lets you plug in your own tools, and credential refresh across interactions keeps long sessions alive when tokens expire mid-task.

DoorDash's benchmark exposes gaps in AI code review: The delivery giant just shared DashBench, a benchmark that replays its own historical PRs to test whether AI reviewers catch real issues or just write comments that sound right. In a twist, a Kimi K2.6 and Claude Fable 5 combo beat DoorDash's own production setup on weighted recall. The takeaway for engineering teams is that no single model dominates, so measure the tradeoffs yourself.

Debugging mobile apps is weird. Shoutout to intermittent connections, mid-onboarding drop-offs, edge cases on devices you've never tested. Unlike most observability tools which never see any of it, bitdrift captures 100% of mobile data, unsampled and in real time, so it’s immediately queryable by engineers and AI agents.

Find and resolve every crash, freeze, slowdown, ANR, and OOM before your users do.

Try bitdrift: mobile observability built for the real world.

A hot take lands. Notion design engineer Geoffrey Litt spoke at the AIE stage last week. After his talk, he posted a 36-tweet thread that engineers can’t stop passing around. He says you should still understand the code that your agents write. But the reason he gives for this is what makes it so interesting.

The obvious answer expires. Ask most engineers why understanding code matters and you'll get one word: verification. You read the code to catch the agent's mistakes. But agents are getting good at catching themselves. If review is the only reason you stay close to the details, you're holding onto a job that's steadily being automated out from under you.

Understand to participate. Litt's actual argument is that understanding is what fuels your next idea. A project is really just dozens of loops with an agent. Your mental model of the system decides how good each next move will be. If you skip that work, cognitive debt starts to pile up. Eventually, you won't even be able to steer your own project.

Borrowed from the classroom. So he built an explain-diff skill that turns every change into a teaching doc. This gives you the intuition before you see the code. Each one ends with a five-question quiz, and he refuses to ship until he passes it. For the full picture, you can read the entire thread. You can also watch his AIE talk when it drops.

Codex is usually limited to its interactive session. This forces you to manually paste in logs or API responses and then copy the results back out. OpenAI has introduced a fix with codex exec. This allows you to run Codex as a standard shell command.

Since it only outputs the final answer, you can pipe it together with other commands like grep or curl. This lets you feed data in and send the results exactly where they need to go.

Runs are read-only by default, so you'll need to include "--sandbox workspace-write" if you want Codex to edit your files.

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

How to build long-running agents for complex workflows (by LangChain): In this tutorial, you’ll learn about Deep Agents, an open-source framework built for complex, long-running AI tasks. You'll also understand how to build multi-agent systems, manage execution environments, and handle context flow.

TryCase: This tool gives each agent its own temporary Linux cloud environment to run the app and test every single change. It also captures screenshots and records video along the way. Instead of just getting code back, you get fully verified results.

CodexBar (by the OpenClaw Creator): A lightweight menu bar app that tracks your AI coding provider limits and shows exactly when they reset.

How to manage an engineering team in the AI era: Traditional engineering workflows are broken. Now that AI handles the coding, the real bottleneck is review and verification. In this guide, Fiona Fung, manager of the Claude Code and Cowork teams, shows you how to ditch outdated processes and focus on shipping a great product rather than just writing code.

The Anthropic team explained how to move beyond basic prompting and start designing loops instead, with four patterns for hands-off workflows.

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