Welcome back. Software architecture used to take years of planning, but now it's happening in a single weekend. On Monday, we looked at how the creator of Bun, which is a Node.js alternative originally written in Zig, used Claude to rewrite almost a million lines of code in Rust. That rewrite has now been officially merged. See how it shipped.

Also: How to deploy a fleet of self-evolving agents, build agent UIs that feel polished, and see an Atlassian engineer's 8-year playbook.

Anthropic billing overhaul triggers developer backlash: The AI lab just unveiled a major change to its paid plans. Starting June 15, Agent SDK calls, GitHub Actions, and third-party apps will tap a new monthly credit pool charged at full API rates, separate from chat and Claude Code. Pro users get $20 in credits, Max (20x) users get $200. Developers are pushing back, with some saying they're shifting heavier workloads to OpenAI. See the impact.

Cursor sets agents loose across the entire codebase: The AI coding startup just shipped cloud development environments that let agents juggle multiple repositories at once, much like engineers operating across microservices. An agent can trace a Slack-reported issue to the affected repos and open PRs in each one. The update also brings Dockerfile-based configs, version history, audit logs, and cached builds that run 70% faster.

Notion opens its workspace to third-party AI agents: The productivity software maker just unveiled a Developer Platform that lets engineering teams deploy custom code, sync live data from any API, and bring agents like Claude Code, Cursor, and Codex straight into the Notion app. New Workers run logic in a secure sandbox, and a CLI ties the system together. It's free through August on Business and Enterprise plans.

Teams burn weeks in manual trial and error building agents, with limited visibility into opps to boost accuracy or reduce spend. And a new model or workload change means re-tuning your agent - a "maintenance tax" that compounds fast. 

AI21 Labs treats this optimization pain as a search problem: they built Maestro to automatically map the full tradeoff surface across models, tools and execution policies for max accuracy within production constraints.

The proof? #1 on BrowseComp-Plus with 95% accuracy, efficiently. Here’s how.

Selling AI to the Fortune 500 still requires a human touch. In just two weeks, three leading AI labs came to the same conclusion. Anthropic launched a $1.5B venture with Blackstone and Goldman Sachs to place engineers in mid-sized firms. OpenAI started its own Deployment Company that same week. Now, Google followed this by hiring hundreds of forward-deployed engineers.

Palantir created the playbook. A Forward-Deployed Engineer (FDE) is a software engineer who works on-site at a customer's office for months at a time. Unlike consultants who focus on presentations, FDEs ship actual production code. Palantir pioneered this role over a decade ago, referring to them internally as "Deltas," and saw a 640% return from their 2022 lows. Now, every major AI lab is trying to replicate that success.

The first bet is a knowledge gap. Customer engineers know their own data and compliance needs, while vendor engineers know the AI architecture. Neither side can win alone. FDEs bridge this gap by writing code directly in the customer’s environment.

The second bet is stickiness. Once an FDE team spends six months integrating Claude or GPT into a firm like Goldman Sachs' compliance workflow, it becomes nearly impossible to remove. This setup transforms into critical infrastructure that secures steady revenue. If your engineering team is bringing FDEs onboard, it carries several implications, and here are the four most significant ones.

Small tasks often get buried in your tracker. They aren't urgent enough to justify opening your editor, yet they still nag at you. One software engineer found that Cursor's Slack integration is a great way to clear that backlog. Once an admin connects Cursor via Dashboard → Integrations, you can just mention @Cursor in any channel:

The agent reads the thread, runs in a sandbox, opens a PR, and pings you when it's done. It's perfect for greenfield tweaks, though messy refactors are still best handled in the editor.

P.S. You can find 50+ AI coding hacks here.

How to use MiniMax Agent: This tutorial shows developers how to use MiniMax Agent as an AI coding assistant. You'll learn how to build and deploy websites, analyze local datasets to generate PDF reports, and explore built-in features like video creation, all while managing projects directly through its file system interface.

Kilo Code: An open-source AI coding agent that works across VS Code, JetBrains, and CLI to help developers ship code faster with agentic workflows. You can generate code, automate reviews, debug issues, and ship faster with AI that truly understands your codebase.

Plannotator (5.3K ⭐): This repo lets you annotate agent plans before they execute and review AI-generated code with a built-in diff viewer. You can collaborate with your team and send feedback directly to the agent to keep your workflow moving.

How OpenAI built the Codex sandbox for Windows: Running Codex on Windows used to force a trade-off between annoying permission prompts and risky full access. OpenAI fixed this by creating a custom sandbox that uses specific user accounts and firewall rules to safely lock down file and network access.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 270K+ 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