Meta ships Muse Code with multi-agent workflows

Welcome back. Coding agents are getting ridiculously fast and capable, but there’s still one annoying bottleneck: our orthodox methods. We are probably not able to train them that well. A senior SpaceXAI engineer just dropped a guide on how to let agents verify and fix their own code. She's using it to ship 2,000 PRs a month. Read details below.

Also: Typing prompts out of thin air (quick demo), a fun gaming project built with AI, and 10 ways to clean AI slop from your codebase.

Cursor engineer shares how she ships 2,000 monthly PRs with agents: Lauren Tan, a senior engineer at SpaceXAI, just dropped part one of her guide on adopting coding agents, and developers are raving about it. The guide already racked up over 500,000 views. She explains how she ships 2,000 PRs a month using self-verifying agents that test and fix their own code to eliminate human bottlenecks.

Meta ships Muse Code with multi-agent workflows: The social media giant just took its agentic coding CLI out of beta. Muse Code is built for complex engineering tasks that a single agent can’t handle. You can now run one command to split jobs across several sub-agents in parallel. These agents also share context and sync state as they work. They also released a TypeScript SDK for building custom agents and moved to monthly plans alongside their existing pay-as-you-go plans. Watch how it works here.

Vercel drops a public file to fix AI slop: The cloud platform just published design.md, a public Markdown file that any coding agent can load to build pages that actually look like your brand. It spells out design choices like color, typography, and spacing, so agents stop churning out generic, cookie-cutter UI. While it uses a stylesheet and testing loops to keep things consistent, the team admits it won't fix messy code or replace real components just yet.

Finest calibrates against your work for the best model.

Finest API serves a cheaper config only when it clears your quality bar.

Finest Code brings the same tech to Claude Code usage credits, so you can keep your existing workflow.

Build with Finest

The tone matters. If you're building something that matters, people will ask for updates. Whether it's your boss, an investor, or another team, they'll want to know how it's going. That makes progress updates a real skill. As AI handles more of the coding, clear communication becomes a manager’s true edge. Slava Akhmechet, an engineering leader at Azure, has a great fix for this. His guide is going viral again for a reason.

The reframe. Akhmechet’s idea is simple: every update is evidence. Each one either builds or dents the case for whether you can be trusted to run the show. Done right, a status email builds your reputation, one message at a time. Here are the rules worth stealing: 

Start here. Clear updates give you quiet leverage that builds with every send, as long as every sentence focuses on real progress. For the writing itself, Addy Osmani's 18 rules for better writing work as both a checklist and an installable agent skill.

A new Managed MCP Server connects Claude Code, Codex, Grok Build, and Devin to MongoDB Atlas, giving coding agents direct access to live operational data.

Explore new MongoDB capabilities for the Agentic Era.

How to build a self-improving AI software factory (by an ex-Oracle Engineer): You’ll learn how to build a self-improving AI software factory. This tutorial teaches how to orchestrate AI coding agents like Claude Code and Codex using control planes, automated triggers, and evaluations to achieve consistent, repeatable, and scalable agentic development workflows.

Visa vulnerability agentic harness (2.4k ⭐): Built on Anthropic's Project Glasswing, this repo is Visa's open-source pipeline for autonomous vulnerability discovery and remediation. It maps attack surfaces, verifies exploits, and validates fixes across eleven stages using models like Claude. LLM-generated fixes are still reviewed by humans before merging.

Ultimate Grok Bot guide for software engineers: SpaceXAI engineer Lingxi Li runs his engineering work through a team of Grok Bots. These bots manage his Cursor cloud agents so he doesn't have to. They watch transcripts, unblock stalls, and finish exactly what he asks for. He shared his full playbook, which includes nightly codebase audits and a Notion-backed system that auto-merges low-risk PRs overnight. 

You ask Claude Code for a date field. The agent installs a picker library, writes a wrapper component, and starts a debate about timezones. Ponytail, an open-source skill shared by this engineer, kills that reflex.

Already sitting on an overgrown diff? Run this to get a delete list with files, line numbers, and simpler replacements:

On the repo's benchmark against a real FastAPI codebase, it cut lines of code by 54% on average. Codex and Cursor setups are in the docs.

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

Claude Code limits are dropping, so Anthropic shared a guide on saving credits. One easy fix is a setting you’ve likely missed.

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