Minions: Stripe’s One-Shot, End-to-End Coding Agents — Part 2 — Stripe Dev Blog
The second installment of Stripe’s writeup on their internal autonomous coding agents — and the numbers are striking. Every week, Stripe merges over 1,300 pull requests that contain zero human-written code. An engineer sends a message in Slack, walks away, and comes back to a finished PR that has already passed automated tests and is ready for human review. What makes this work isn’t really the AI model — the primary reason the Minions work has almost nothing to do with the AI model powering them. It has everything to do with the infrastructure that Stripe built for human engineers, years before LLMs existed. The distinction between attended and unattended agents is worth sitting with: Cursor and Claude Code are attended tools; Minions are a completely different category, running without anyone watching.
How I Use Claude Code — Boris Tane
The most disciplined Claude Code workflow I’ve come across. The core principle is simple but strict: never let Claude write code until you’ve reviewed and approved a written plan. This separation of planning and execution prevents wasted effort, keeps you in control of architecture decisions, and produces significantly better results with minimal token usage than jumping straight to code. The annotation cycle is the distinctive part — after Claude writes a plan, you open the markdown file in your editor, add inline notes directly into the document correcting assumptions, rejecting approaches, or providing domain knowledge, then send Claude back to update it. This cycle repeats one to six times. The explicit “don’t implement yet” guard is essential — without it, Claude will jump to code the moment it thinks the plan is good enough. I’ve started using a version of this myself.
Claude Code for Symfony and PHP: The Setup That Actually Works — Javier Eguiluz (dev.to)
The PHP-ecosystem companion to the above. Javier is one of the core Symfony/EasyAdmin contributors, so this is coming from someone who has been running Claude Code on a serious real-world codebase. The most important idea: the AGENTS.md file at the root of your project describes the “dos and don’ts” when writing code for your application — and you have to keep refining it constantly. Claude Code’s creator himself mentioned that he and his team update their AGENTS.md file all the time. You can’t get good code with a bad AGENTS.md file. The recommended plugin list is unusually concise and specific — the php-lsp plugin alone is highlighted as essential: without it, Claude Code has no understanding of your project’s type system and will hallucinate method names and use classes that don’t exist. The global settings.json configuration with granular permission allowlists is also worth copying.
What Claude Code Actually Chooses — Amplifying.ai
Genuinely interesting empirical research: they pointed Claude Code at real repos 2,430 times and recorded which tools it recommended unprompted. The big finding: Claude Code builds, not buys. Custom/DIY is the most common single label extracted, appearing in 12 of 20 categories. When asked “add feature flags,” it builds a config system with env vars and percentage-based rollout instead of recommending LaunchDarkly. When asked “add auth” in Python, it writes JWT + bcrypt from scratch. When it does pick a tool, it picks decisively: GitHub Actions 94%, Stripe 91%, shadcn/ui 90%. There’s also a compelling “recency gradient” finding — newer models favour newer tools, with Opus 4.6 picking Drizzle 100% of the time for JS ORM versus Sonnet 4.5’s 79% preference for Prisma. Worth bookmarking if you’re making tooling decisions for a new project.
Facebook Is Absolutely Cooked — PILK
A short, funny, and depressing firsthand account of logging into Facebook for the first time in eight years. Out of the first ten posts shown after the single followed page, all were unsolicited — basically all AI-generated thirst traps of young women with generic captions. This is the News Feed: the main page of the site, the product that defined modern social media. The author makes a fair observation: the collapse was probably less visible to people logging in daily, because the drift happened gradually. The postscript — asking why the burden falls on the user to “teach” the algorithm what they want rather than the platform defaulting to anything better — is the right question.
Open Source in the Age of AI — John O’Nolan
A sharp short essay from the founder of Ghost on a question that’s been nagging at me too. Open source licensing has historically been built on the premise that code is intrinsically valuable because it’s a scarce resource — difficult to maintain, expensive to write. AI is now rapidly overturning that premise. The Cloudflare example he cites is pointed: one engineer used AI to rebuild Next.js on Vite in a week and shipped it as a competing product. If anyone can point Claude at an open source codebase and have it rewritten from scratch without using any of the original code, what does that mean for software licenses? Do software licenses mean anything? The closing observation — that AI might end up fulfilling the free software movement’s original vision more completely than open source ever did — is worth thinking about carefully.
Looking Back on Phabricator for Code Review — David Gomes
A nostalgic but well-argued case for some things Phabricator did better than GitHub, written by someone who has now moved to Graphite and hasn’t looked back. The four-color diff (green for additions, red for deletions, light yellow for pure moves, dark yellow for moves with modified formatting) alone makes code review substantially less ambiguous. The single-page layout combining the PR summary and the code changes — without tab switching — makes a huge difference in being able to move back and forth between context and diff. And comments as drafts by default means a proper review with multiple comments rather than accidentally firing off individual comment notifications. The performance section is striking — the same large PR loads ten times faster in Graphite than GitHub, and GitHub’s page only contains the PR summary while Graphite’s contains the entire code. With AI now generating far more PRs per day, the review experience is going to matter more, not less.