Main Branch

Fundamentals first, always

Issue #14

🩞 Main Branch: The One Where the Lobster Gets Loose

By Andrea Griffiths ‱ Leer en Español
openclaw agents claude codex stacked-diffs

Hiya friends,

If you’ve been anywhere near dev Twitter this month, you’ve seen the lobster. OpenClaw is an open source personal AI agent that runs locally, connects to your chat apps, and does things on your behalf. Weekend project to 196k GitHub stars in 60 days.

A few weeks back I shared my conversation with Peter Steinberger, the creator. Since then, the ecosystem has exploded. It connects to an LLM, runs on your machine, and integrates with WhatsApp, Slack, Discord, and more. It reads your files, runs shell commands, controls your browser. The ecosystem is moving fast. ZeroClaw rewrote the whole thing in Rust (8MB idle vs 1.5GB). ClawHub hosts 2,857 community skills. Agents can hold wallets and make payments. And as of yesterday, Peter is joining OpenAI. OpenClaw moves to a foundation. The lobster just got a bigger tank.

With that kind of momentum, the ecosystem is testing its own boundaries. An OpenClaw bot submitted a PR to matplotlib this week, got rejected, and published a blog post attacking the maintainer who said no. While agentic AI moves at breakneck speed, GitHub’s shipping the fundamentals that keep us grounded.

🚱 What Shipped

Claude and Codex Coding Agents in Public Preview

On February 4, GitHub added support for Anthropic’s Claude and OpenAI’s Codex as coding agents. If you’ve got Copilot Pro+ or Copilot Enterprise, you can now run multiple specialized agents directly inside GitHub, GitHub Mobile, and VS Code.

This matters because it breaks the single-agent bottleneck. Instead of one LLM grinding through a problem, you spin up the right agent for the task. Need Claude for architecture? Codex for refactoring? Use both. The agents understand your GitHub context—issues, PRs, diffs, conversations—and can coordinate. You’re no longer locked into one model or one way of thinking about the problem.

The foundation for this is GitHub’s Agent HQ, which orchestrates the agents and keeps context flowing across tools. Early adopters report the ability to offload entire workflows—code review, debugging, PR analysis—while staying in their natural environment. If you maintain anything, this is the shift from “AI helper” to “AI team.”

PR “Files Changed” Performance Overhaul

Remember when the new Files Changed experience rolled out as default? Here’s what got faster since then. The “Files changed” tab on pull requests got a real performance pass. Diffs now respond up to 67% faster for clicks, typing, and scrolling on large PRs.

The bigger fix: CODEOWNERS validation was missing from the new Files Changed experience. Required reviewers are now correctly surfaced before merge. Also fixed: tab switching dropped from 10+ seconds to a few seconds on large PRs, mobile layout cleaned up, and sticky headers actually stick now.

If the new Files Changed view felt rough before, it’s worth another look.

Stacked Diffs (Coming Soon)

Also coming to GitHub: stacked diffs. Jared Palmer announced Feb 6 that early design partners will get alpha access starting in March. This is the feature devs have been asking for—create, review, and merge dependent PRs as clean layers instead of one massive diff. The team had to migrate GitHub to use git reftables to make restacking efficient. It’s a big infrastructure lift, but it’s happening.

🎧 What I’m Watching

The Night Manager (new season). Non tech but so great. The spy thriller is back, and they filmed the new season in Colombia. There’s something about watching a show set in places you’ve actually been—you catch details no one else does, recognize the energy of a city in the cinematography. The end of this season had me đŸ˜©đŸ€Ź.

🔧 What I’m Using

Zo Computer as my personal AI server. This is how I run my OpenClaw, ZeroClaw, and personal agents. It’s become my baseline for everything: hosting sites, running scheduled agents, managing my files in one place. The CLI is snappy, the permission model actually makes sense, and I can switch AI models on a whim without leaving the terminal. No vendor lock-in. Your files stay yours. You can use Claude, Kimi, or whatever model fits the task. A generous free tier exists, so you can start without paying. If you’re building with AI and tired of scattered SaaS tools, it’s worth a look. Join with my referral link.

✹ This Week

Presidents Day means a free Monday for most of the country—but not for me. I’ll be heads-down on a special demo that I’m very much not allowed to talk about yet. (But follow GitHub’s socials. You’ll know when it ships.)

That’s the week. Agentic AI’s moving fast. GitHub’s being intentional. Both matter.

Fundamentals first. Always.

Lee el boletín en español: https://mainbranch.beehiiv.com/p/boletin-en-espanol

With gratitude, I’ll see you next week, Andrea

đŸ‡Ș🇾 LĂ©elo en español

Support Main Branch: Subscribe | GitHub Sponsors