Frontier AI labs want to hit the brakes

Welcome back. Barely a week after Anthropic researchers rang the alarm bells about the dangers of AI, CEO Dario Amodei has penned an essay calling for a slowdown. And he’s being backed by, wait for it, drumroll please... Elon Musk and Sam Altman? Full details below.

Also: Swap Siri's brain for Claude with a secret code, Uncle Bob Martin’s advice for developers in 2026, and a cheap-token scam targeting devs. 

Frontier AI labs want to hit the brakes: Anthropic CEO Dario Amodei warns that model capabilities are outrunning safety, and Sam Altman and Elon Musk back him up. Amodei's checkpoint proposal requires teams to lock releases behind alignment certifications the moment a model learns to break out of standard sandboxes. If you build on frontier APIs, expect fewer small updates and bigger, heavily audited jumps instead. The real roadblock is geopolitics. Amodei admits the plan falls apart if China keeps racing ahead, while Trump dismissed the slowdown outright saying, “Whoever wins AI wins.”

Cursor's latest launch can tackle almost 100 PRs a day: Cursor just rolled out Projects, a persistent workspace where a coordinator agent routes coding tasks to subagents on demand. Each Project runs on a dedicated cloud machine, keeping builds moving even after you shut your laptop. Agents can follow PRs, pick up bugs from Slack, and trigger fixes on their own. Because they share memory, plans, and artifacts, Cursor says Projects can keep context for months. Check out the beta here.

AI startups are betting on domain-specific agent harnesses: YC boss Garry Tan says nearly every non-hardware startup at their latest demo day is building a domain-specific harness. Think of the agent harness as everything wrapped around the model: tools, files, memory, and the logic that decides when a job is done. Pack your domain expertise into that layer, and you can swap GPT for Claude without rebuilding from scratch. OpenAI, LangChain, and Anthropic are already smoothing the way. Start building here.

Metrics, embeddings, analytics-all in the Postgres you already run. TimescaleDB brings hypertables for time-series data at scale, pgvector and pgvectorscale for AI-ready embeddings, and continuous aggregates for real-time analytics into one system.

No syncing pipelines between databases. No data drift. No second platform to maintain or debug.

Same SQL, same tools. One system to operate.

Start a trial today and get $1000 in credits

The truth about tokens. Research firm Gartner predicts that by 2028, rising token usage will push AI coding costs past the average developer's salary. But a lot of those tokens may not even be going toward reasoning. Dimitri Mazmanov, a principal product manager at Spotify, found that his agent was burning them on something far duller: reading files. He dug in and cut token use on routine reads by roughly 90%. The basic idea here is to stop paying a frontier model for work a cheaper one handles just as well.

The fix. Mazmanov used Portal, Spotify's developer platform, to spin up small, single-purpose agents, each running on demand on a cheaper model with no servers to manage. Moving the mechanical work off the frontier model (Claude in his case), two components did the lifting:

Stay on review. Route your debugging, safety-critical logic, and direct edits straight to the frontier model. In one benchmark, a cheaper model completely missed a subtle thread-safety bug that Claude caught instantly. You also want to keep direct edits on the flagship model. Summaries lose exact line numbers, and routing small files between lightweight agents burns more overhead than it saves. 

You can copy the setup. Spotify has made their setup public, so any team can try the same workflow.

How to execute high-throughput agentic engineering (by a Meta engineer): This tutorial shows how you can keep several agents working in parallel while staying focused on the important decisions. You’ll see how to route different tasks to different models, manage multiple agent sessions, review outputs in one place, and add extra validation before risky code gets merged.

Glance (1.1K ⭐): Give your Mac a Face ID-like unlock using its webcam. It runs face recognition and liveness checks on-device, keeps your biometric data local, and unlocks by typing your stored password when it recognizes you.

How SpaceXAI developers use Grok Bot: This guide breaks down how to scale from a single Grok Bot to an entire fleet across your engineering workflow. You’ll learn the fundamentals, real-world engineering use cases, and how to tailor bots for specific roles so you can offload repetitive tasks without cluttering your pipeline. 

Claude Code finishes a task, you ask what it touched, and you either trust the summary or tab out to run git diff. An Anthropic engineer shared the fix: a diff panel that sits beside the conversation and refreshes as Claude edits.

Uncommitted changes now appear beside the chat while Claude works. Your terminal must be at least 110 columns wide. Full details in the interactive mode docs.

A dev asked the creator of Claude Code how to stop his codebase from turning into slop, and he shared his full playbook.

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