Agent-native Architectures: How to Build Apps After the End of Code

Claude Code's co-founder just dropped a banger plugin: a code simplifier they internally use at Claude/Anthropic, for all vibe coders to benefit just type in session: /plugin install code-simplifier i'd definitely use it to have a better codebase that can scale tbh

onboarding was painful so i made a native mac app thanks to cloudflare it now: • automatically starts a local bridge on your mac • spins up a secure tunnel (no ports to open) • deploys an AI agent to a worker running locally • hands you a URL that works from anywhere https://t.co/4Rbimr1DDk

Introducing my latest tool designed to accelerate agent coding workflows: repo_updater (ru for short). You can get it here, as always for free and 100% MIT-licensed open-source: https://t.co/nONU9xSlT8 I basically made this tool out of necessity, because I was wasting far too much time and energy managing an ever-increasing number of public and private GitHub repos across 4 different machines (a Mac at home, a Linux workstation at home, and two remote bare-metal Linux servers in the cloud) that were always drifting apart and causing mental friction and wasted time. It also caused low-grade anxiety because I worried about agents doing silly things and wiping out useful work. Sometimes I would accidentally work on the wrong machine (this happens much less often now that I have each host automatically displayed in a different color scheme in Ghostty and WezTerm, see my recent post about that or check my misc_coding_agent_tips_and_scripts repo). So just this simple workflow of pulling remote changes and pushing local changes and making sure repos are in sync in a smart way, but for a large list of public and private repos and done in a cross-platform way (the system is pure bash scripting that works with the gh utility from GitHub) that is parallelized across repos, was a big unlock for me in terms of automation. But of course, I'm not even the one using ru, although I certainly could; my agents are the ones that use it on my behalf. So, naturally I designed ru so that it's "agent-first" in every way, ensuring that it was as ergonomic and intuitive as possible for use by coding agents, for which I had them design their "dream tool" in an iterative process using the robot-mode prompt I shared earlier (i.e., prompt number 3 in the "My Favorite Prompts" series, lol). So how do I use it exactly? Well, step one is to install it, which takes 2 seconds using the curl | bash one-liner script given in the readme file of the ru repo. Then, suppose you would like all your repos to live in the /data/projects directory on your Mac or Linux machine. The next step is to get a list of all of your public and private repos you want to manage with ru. To do that, you can create a text file with one URL per line or use several other formats. One particularly easy/lazy way is to simply ask Claude Code to use the gh tool to list all your repos and then specify it from there (e.g., "Take all the non-forked repos with more than 3 files in them that I touched at least once in the past 3 months and add the public repos to the public repo list of ru and the private repos to the private repo list of ru.") Then, simply start up Claude Code and use this prompt: "First, I want you to cd to /data/projects and then run the `ru` command just like that; then in the same directory I want you to run the command `ru sync` and carefully study the resulting output. When you've done that all meticulously, reviewing carefully every single line of output, I want you to help me to ensure that all my repos are up-to-date; BUT, I want to be super careful and hyper cautious and vigilant about potentially losing ANY useful work (code, documentation, beads tasks, etc.) in BOTH the local repos and the remote repo on GitHub. In every case, I want the one "best" canonical version of each committed file (usually that is the latest version, but NOT always); this cannot be mechanically determined and must be done by you manually and carefully by doing diffs between versions and understanding the changes and differences and what they mean relative to the purpose and structure of each specific project. Also, use common sense to avoid committing obviously ephemeral files or build artifacts, sensitive .env files containing credentials, etc. Use ultrathink." Do this on each of your machines. The best way is to do it first on your main machine, and then have Claude SSH directly into your other machines and install ru on them and configure them identically in every respect (to maintain sanity, I highly recommend using the same directory for all repos, usually `~/projects` or `/data/projects` for me). Then Claude can run the same process on the remote machines, or you can directly connect to them yourself and start up new Claude Code sessions (this is what I recommend) because these can get very interactive when you have complex situations that need your guidance and feedback. To see some examples of these complex interactive sessions using CC and ru, see the attached screenshots. Anyway, this project is already saving me a ton of time and mental energy, and I didn't even discuss the more powerful functionality in there for helping to automate the management of GitHub issues and PRs, which is the main time savings and automation potential this tool provides me with. That will have to wait for another post in the next day or so.

I have a slightly different take on the Anthropic situation......... People are abusing Claude Max plans and running them automated (Ralph Wiggum etc) That's NOT what these ((very, very generous)) plans were for and it says so in the TOS. Adjusted some of my own projects but I get why: you're getting like $1500 worth of API usage for $200 on Max. They're subsidizing this for HUMAN users. Not for Ralph Wiggums. I get that there's frustration but throwing fits will probably get Max ended. You do not want Max ended. Isn't Anthropic's POV on this reasonable?

A ton of people have asked me in the past week to record a screencast because they want to watch me work and understand my workflows better. I'm not willing to do that because I know it will kill my velocity, since I'm constantly jumping back and forth between public and private projects to push all my initiatives forward every single day. And I also don't want to be stressed out worrying that I leaked .env details or inadvertently violated an NDA with one of my consulting clients. But I AM willing to do the next best thing, which is to build stuff live, in public. Normally this would be about as exciting as watching paint dry, but I'm going to make this project RIGHT NOW in a single day. This started out as a dumb joke this morning and a domain purchase on a whim but has already blossomed into an almost unbelievably ambitious and full-featured web app and CLI tool for humans and agents, as well as a "how to" site and hub for agent skills files. This is going to be the home for all my prompts, both individually and as collections of related prompts, and also associated skills. The skills will relate to generic prompting workflows and also to tooling, both my own agent flywheel tooling and other dev tooling (e.g., gh, gcloud, vault, wrangler, vercel-cli, supabase-cli, etc.) In what is surely not the first extremely "meta" situation here, part of the new project is itself a nice "How I conceived, designed, planned, and implemented all of this in one single day" rich webpage that basically will provide the complete Claude Code session (I'm keeping it all in one mega session for continuity) that will be beautifully presented along with explanatory narrative that shows the thinking behind each prompt and workflow. You've only "missed" the first part of the process, where I create the initial markdown plan document (I put 'missed' in quotes because later you'll be able to see all this at the session level on that "/how_i_made_it" page). But I'm not even done with the markdown plan; I'm still doing the final polishing and checks using Claude Code, and then I'm going to do a couple of passes using GPT Pro to improve it as I've explained in my recent posts. You can see the current version of the plan right now, which is already extremely good and complete (it's 6,452 lines already, lol): https://t.co/xi3qWQ6U6v Then after I revise the markdown plan, I'll turn it into a ton (probably at least 300+) self-contained beads. Once that's done, I can use my bv tool to export those beads to a static GitHub Pages site so you can explore them and understand how markdown plans map to beads in my system. Then I will iteratively polish and improve the beads until the boys can't come up with anything left to improve. Then I'll boot up ye Old Agent Swarme and knock this thing out in hours for you. If all goes according to plan, I should have the live site ready for your enjoyment late tonight or, worst case, tomorrow.

INTRODUCING: @doanythingapp Do Anything Agents are a totally new kind of agent They: - work independently for weeks/months+ - have their own email - self manage entire projects - can use almost any tool on the web Today the alpha opens to the public. Here's how they work: https://t.co/PUMYrMBGxJ

Claude Codes

If you're building anything with software right now, you need to read this guide (and sub to every). If you're not building agent-native software now, you're gonna be completely re-building your product in the near future.

If you're remotely technical and remotely good at getting in front of decision makers at mid-market companies that do a few hundred million of revenue and need to increase their margin profile, please just go into AI consulting. You'll bootstrap to 100s of Ms if you stick with it https://t.co/uHgjVhSLHC

the 2026 ai engineer roadmap

some of anthropic rugpulls since August 2025: > revoking Max users access outside of Claude Code & banning paid users > cutting off xai api access > 5 years retention of all conversations and code, all data will be used for training > 1.58-bit quantized models during daytime/nerfed models > plus not getting opus 4 in claude code > max plans limits cut in half 6 weeks ago, no comms > weekly limits without concrete numbers > 5x/20x plans being actually 3x/8x of plus > DMCA takedowns of repos that have to do with Claude Code, one of which was my own personally > windsurf no access to claude 4 > cutting off openai api access what an absolutely horrible company

Wrote a quick explainer for Plan Mode, and why it's fundamental to making AI Coding actually work. A new world is a shift-tab away

I Built 3 Claude Code Commands to 10x My PM Productivity

We just open sourced the code-simplifier agent we use on the Claude Code team. Try it: claude plugin install code-simplifier Or from within a session: /plugin marketplace update claude-plugins-official /plugin install code-simplifier Ask Claude to use the code simplifier agent at the end of a long coding session, or to clean up complex PRs. Let us know what you think!

Claude Code is changing how The Browser Company hires and works

Two great engineers quietly break down the mindset shift required to survive in the AI era https://t.co/XFGy3Uzv7M