Meta ships a personal agent that writes and runs code

Welcome back. One prompt to one model no longer cuts it. OpenAI says it pointed 10,000 agents at one problem and cracked a $1M math puzzle in about 88 hours, only to spark a controversy over who got there first. Meanwhile, Meta dropped Muse, a single agent on its own cloud computer that works after you close the app.

Also: A code to help fix your back pain (no seriously), register for 3 days of GrokBot tutorials by SpaceXAI, and watch how a dev turned his Kindle into a terminal.

Meta just dropped an AI agent that handles your chores and writes code: The social media giant just dropped Muse, a personal AI agent. It handles background tasks, browses the web, and executes code all on its own inside a sealed sandbox that keeps untrusted web data away from the core execution layer. A separate Sentinel agent must approve every outbound action, and it asks for your approval before any big moves. It seems like Meta is competing against GrokBot on the security angle. An agent this powerful needs trust, something Meta has struggled with in the past, but they promise it's fixed this time. You can try it here.

Did OpenAI really crack a Millennium math problem? The ChatGPT maker says its model just produced a 100-page proof for Navier-Stokes, a problem that’s one of mathematics' seven unsolved Millennium Prize Problems. But there’s more to it. NYU mathematician Tristan Buckmaster says his team had already fed related unpublished drafts into OpenAI’s Codex, and the AI lab won’t say if those drafts ended up in the training data. The feud got even dirty when OpenAI allegedly pressured Buckmaster to publish without his co-author, who works at Anthropic, warning that going public could wreck his career. 

Anthropic researcher quits, warns AI could hack anything: Jacob Coxon, who spent three years on pre-training research at both Anthropic and OpenAI, claims both labs are racing toward super-intelligence. Which is basically an AI that upgrades itself with almost zero human oversight. He warns these models could eventually hack anything and reshape the world overnight. Anthropic Alignment lead Evan Hubinger agrees, putting the odds of “AI ending humanity” in the next decade at over 10%.

Growing Postgres databases eventually hit a wall: queries slow, dashboards lag, and adding another database starts to feel inevitable.

Tiger Data keeps Postgres fast as your data grows with time-series primitives, columnar storage, and automatic partitioning. Same SQL and tooling. No pipelines or second system.

Just Postgres, built to scale with your workload.

Start building for free today. No credit card needed.

It shouldn't be possible. Agents usually get more expensive as more engineers use them. When costs spike, most companies either cap usage or hunt for discounts. Uber found a better way. Between February and August, agent requests surged 9.4x, and users grew 7x, yet their AI spend stayed flat after April. For one specific model, the cost per session actually dropped 52% from its June peak. 

The bill is an engineering problem. Rather than looking at one massive number, the team broke down its spending into the core drivers: user adoption, workload, and token pricing. Since adoption was supposed to grow and they couldn't control market prices, they focused on the only lever they had: making every agent more efficient. 

Uber tackled that from a few different angles. Here were the biggest wins: 

The takeaway. Your AI bill is just another system to optimize. Uber didn’t cap engineer use or lean on discounts. Instead, they reworked how the agents run to absorb the growth.

Grok Bot is your team of AI teammates that get work done. Learn how to get the most out of Grok Bot at this three day event.

See live demos featuring the team behind the product and discover best practices for customizing your bots based on your role and department.

Register now

How to build agents that create their own UI: In this tutorial, CopilotKit’s Tyler Slaton shows three ways to build great UIs for AI agents. You’ll learn how agents can create dynamic frontends, handle user input, and power interactive workflows like incident triage apps.

Diagram Design (35.4k ⭐): You can turn your codebase, database schemas, or architecture into easy diagrams instantly. This tool gives Claude Code, Codex, and other agents 39 professional styles, from flowcharts to user journeys, that match your brand’s colors and fonts.

How to cut Claude API costs without hurting performance (by Anthropic): This guide shows how to lower Claude API spend with better prompt caching, cleaner instructions, and smarter effort settings. It also walks through tools like “prompt-audit” and “cost-optimize”, which can catch wasteful prompts, trim token use, and test cheaper configurations without tanking performance.

Long /goal runs can drain your weekly Codex quota before you check back in. This dev shared a fix: GPT-6 Astra can see your remaining usage % and respect a budget written in plain English.

Start a /goal in Codex with Astra selected. Add one budget line to the prompt:

Astra checks usage as it works and stops at the floor, so overnight runs no longer eat the whole week's quota.

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

Spotify's engineers cut Claude Code token usage by 90% with a two-model routing setup, and their breakdown explains exactly how.

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.

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