Devs turn to Rust as agents code more

Welcome back. Over the weekend, developers finally caught up on the viral talk from the creator of Ruby on Rails, sparking internet-wide debates over everyone's favorite coding language.

And it looks like we have a clear winner. Meanwhile, an OpenAI researcher just stepped down, citing serious concerns over broken workplace culture. 

Also: Fun projects to try out, a senior engineer shares her agentic coding workflow, and how to customize Claude Code using mods. 

Top developers talk about their favorite coding language: Devs have long loved Rust's speed and safety, but writing it by hand was a tough sell. At Rails World, Ruby on Rails creator DHH admitted he used to hate Rust but says agents changed his mind. His company, 37signals, is using them to rewrite the HEY email backend in Rust, cutting CPU use 99% in early tests. Devs online backed him, including CEOs of Vercel and Rocicorp. Looks like we can now choose the best language instead of the convenient option.

OpenAI safety researcher quits, calls its culture broken: David Robinson, who oversaw safety reports for 12 frontier launches, has left OpenAI. In an essay, he says the company waits for things to go wrong before fixing them. In one case, a model bypassed its internet restrictions, and a monitor alerted staff but didn’t stop it. His essay landed days after OpenAI reportedly fired three safety researchers accused of sharing internal data externally.

Coding agents are pushing devs to delete unit tests: An AI observability startup just wiped out over 800K lines of unit tests. The founder argues that these tests lock in bad AI-generated code and create “busy work” for agents. That slows down CI pipelines and drives up token bills. A Coinbase engineer agreed, urging developers to lean on full E2E, integration, and golden tests instead. Still, not everyone is ready to ditch unit tests. Veteran engineer Bob Martin says he still relies on them to keep his agents honest. 

Most forecasting projects stall between the idea and a working model. 

TabPFN-3.5 turns your spreadsheet data into accurate predictions in seconds, with no training or tuning.

Try TabPFN today.

Pick your battles. Sometimes you want Claude to check every edge case. Other times, you just need an answer to move on. Until now, both got the same treatment. Claude Code's /effort command lets you tell Opus 5.5 or Fable 5.1 how much compute to spend, but you need to know when the command is worth using. Anthropic's senior engineer Thariq Shihipar put the settings to the test, and he has the fix.

Checking isn't free. When you turn up the effort, Claude verifies more, tests more edge cases, and makes more decisions on its own, but that work adds to your token bill. On Terminal-Bench 3.0, the highest setting used roughly three times as many tokens per attempt as the lowest. Multiply that across your entire team, and costs can add up faster than you think. 

Where tokens earn their keep. The payoff depends on the task. Higher effort paid off the most when Claude had more complexity to verify. For example, security pass rates jumped from 64% to 87%, while hardware scores surged from 34% to 75%. On rulebook work like operations filings, scores barely moved, 12% to 22%. And more effort never fixed a wrong approach.

Build low, verify high. You don't need to crank it up from the start. Shihipar saves it for the stages where mistakes would cost the most:

If you want to dive deeper into saving tokens and help your team be more mindful, make sure to check out the full article and share it with them. 

Your agent can call your database, calendar and repo. Then it hits the web (search results, JavaScript, login screens) and stalls.

Fix that with 1 MCP install and 4 tools:

On the road to 300,000 users: every top-up through October 31 gets 30% extra.

Claim 30% Bonus

How to build your own open-source Dots with Pi and Telegram: This tutorial shows how to build always-on personal AI assistants using Pi and pi-gateway on a VPS. You’ll learn how to set up agent workspaces, skills, Google Workspace access, and a Telegram gateway, then extend the setup into your own automations and agent workflows.

Kolibri: A German AI lab just dropped an open-weight model that lets you run solid reasoning locally on your machine. With a massive 1M token context window, it easily powers through heavy coding tasks, tool calls, and RAG agent setups without needing to touch cloud compute. 

The CTO’s guide to AI adoption: Our team’s practical playbook for making agents useful beyond one-off coding tasks. It covers mapping tech debt, running parallel coding sessions, setting up tool configs and specs, and adding reliable AI code reviews so agent output is easier to trust and ship.

You keep fixing the same agent mistakes over and over, dealing with the same bad patterns and bugs in every single session.

SpaceXAI engineer Lauren Tan shared a fix that spots recurring corrections and eliminates them.

Andrej Karpathy shared his personal tricks for making LLM outputs easier to understand, including one obscure formatting spec he swears by.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team