Dylan Mikus

Software production is industrializing. Build your software factory. Use any code agent you want. We're making infra to help you agentify your software development lifecycle

@adamhjk Gonna take a while for the industry to catch up, but I think we're seeing software engineering industrialize Think we'll have to take inspiration from manufacturing, industrial engineering, and operations research

That’s why it breaks the SDLC - you can’t review it all, because it happens too fast. You can’t only improve the quality of the machine that catches issues earlier, and only after they’ve arrived. It’s wild.

Yes, if you’re trying to use AI in the context of your existing SDLC and organizational structure. Absolutely not if you’re building the machine that builds the machine - in that phase, the accelerated refactoring loop lets you find and repair the seams in the software

@shivsakhuja Congrats on the public launch! Very cool product, onboarding was really streamlined and I loved the goose avatars

Introducing Gooseworks — AI coworkers that can do real GTM work. Over the last few months, we've been obsessed with turning AI into actual coworkers for our team. And it's working. Goose is now our most valuable teammate – it finds high-intent leads, runs outbound campaigns, coordinates with influencers, tracks SEO / AEO and a ton more. Today we're launching this publicly so everyone can create their own AI coworkers. Each coworker (or Goose) has its own computer, filesystem, mailbox, memory, and tools. Goose can work autonomously for hours, communicates over Slack, iMessage, WhatsApp or email, and gets smarter as you give it more context. It's not just a personal assistant — your whole team can talk to Goose from anywhere, and schedule it to run on autopilot. And here's the best part – we've already given Goose 100+ skills and data APIs for GTM work: – map your TAM – search people databases with natural language – find leads from X and LinkedIn posts – monitor intent signals from your ICPs – update your CRM and generate reports – find influencers and run outreach campaigns – track visibility in search and answer engines – connect to all the SaaS tools your business runs on +++ lots more. We also open-sourced every skill as a toolkit you can install directly in Claude Code. Comment "Goose" and follow and I'll send you the install link.

self-reviewing my vibe coded PRs like

@edsim Switch to codex subscription?

It was great while it lasted but simply not sustainable for my OpenClaw bot. Waiting for my Mac Studio to run my own open source models

@KentonVarda Gonna be choppy, but overall good

Honestly "AI that can find every vulnerability" sounds way better for the good guys than the bad guys. Not sure why everyone is losing their minds here.

@samagra_sharma Thanks for the summary! I would not have made it through all 244 pages 😅

I'm gonna crash out the next time I see another "first AI agent to do XYZ"

@garrytan Whole computer needs to be versioned and the agent needs to be able to read back and undo its changes Like git, but for the self-modifying AI computer

OpenClaw is amazing until it does something funny like alter its own config file because you asked to try the new dream mode and then it can’t restart because of invalid config with the wrong version. This happened to me yesterday morning. I had to go ssh in with Claude Code to get it fixed. But the sheer fact all of these can happen shows you we are at Apple I stage: assemble your own motherboard stage. But the Apple II personal computer moment (where anyone can go to the store and buy a PC and it works) is coming for OpenClaw We are so early and this is the worst it will ever be

@badlogicgames What are your thoughts on how to make money from making building blocks?

recommended reading.

In general, I think "worse is better" and the Unix dev philosophy wins Software that's easiest and quickest to use wins over software that's better, because it has better evolutionary advantages to adoption and contribution AI might change it, because implementation is now cheap

@NathanFlurry @appfactory Think isolates make sense for use cases that mostly need file systems and JS code gen Think it's too great an uphill battle for general purpose computer agents

Earlier this morning I said "my sandbox is a devcontainer, what else do I need?" It sucks because it's slow, but it works! This video is convincing. I know Theo didn't point to particular piece of tech, but I think isolates are probably the contender. It does mean that we need to bridge almost everything that cannot be run in a isolate via a TypeScript interface (e.g: Postgres, MySQL, Redis), but that's totally doable. My mental model for this is basically something like React-Native; but made for agents: "Compute-Native?"

@calvinchen Yes, Claude's computer use via Chrome integration is crazy slow. I have a workflow that runs out of "concurrent tool calls" all the time and I need to just type "continue" partway through.

i tried Claude Cowork the other day for manual tasks on platforms like Deel and Ashby my takeaway was that computer use feels best when the task is tightly scoped and visually explicit. if i gave clear instructions and screenshots for navigation, it could do okay. once the task became more ambiguous, it seemed to struggle much more, to the point where i just took over what stood out even more, was how slow and clunky the workflow felt. maybe part of that was my setup, since i was using the claude desktop app with chrome integration, but having to jump through multiple hoops just to get started added a lot of friction (i assume this will be fixed soon) i still much prefer working from the terminal and asking for a CLI tool or MCP integration for whatever platform i’m using. there’s less context switching, the work feels more abstracted away, and agents seem much more comfortable operating under ambiguous directions when the domain is code or text rather than a browser UI feels like computer use is useful for narrow, well-specified ui workflows, but MCP/CLI integrations are still a much better interface for more open-ended work for agents

@andrewqu @theo CLI first is the way 🧙‍♂️

Anyone building a new agent orchestrator like t3codes, superset, etc Please bundle in a CLI to make it easy for me to "import" a project rather than needing to open up the UI, file picker, and go 1 by 1 ▲ npx t3 add ~/projects/some/folder and auto open afterwards @theo

@0interestrates I hope Michael Novati is still at Meta and asked for PR changes

it's easy to approve zuck's diff, but do you have the courage to request changes on zuck's diff?

@hausdorff_space @CoastalFuturist @pierrecomputer Feel like if it's hard for people to build and little details affect the performance and correctness, then it's hard for infra to build As of now, I'd never trust a vibe coded infra, security, or auth product

In contrast, @CoastalFuturist was telling me that he had met multiple people who attempted to vibe code @pierrecomputer’s git storage server, and then later came back to buy it. Hilarious. It is very hard to explain which things are going to be difficult and which aren’t.

@mitsuhiko Can you nest tasks or steps?

Absurd (next to pi) is probably the infrastructure decision that made me the happiest so far. It's so nice and easy to use. I did some updates on the docs today which were pretty atrocious before.

@gakonst @dwr 👋 lets build some stuff

looking to hack with ~15-20 friends in person in NYC the week of april 14th, plz dm me if you're a cracked dev & wanna hack together in person, i'll provide office pizzas & diet coke

@gm_mertd @ssh_exe_dev @amikadev Cool! I'll try it out

@dbmikus @ssh_exe_dev @amikadev runruntime.com

Whoever builds the simplest (repeating: simplest) sandbox infrastructure will win that market. There is literally no company doing this right now. Before you drop a reply here: yes I tried it. No you are not. Closest I have seen is @ssh_exe_dev but it is not really sandbox provider for platforms Kudos @davidcrawshaw

@gm_mertd @ssh_exe_dev @amikadev Sorry, I guess I don't understand. What are you using sandboxes for? Are you building a platform / product or are you using them to sandbox coding or general purpose agents for your own dev work, etc?

@dbmikus @ssh_exe_dev @amikadev yep exe.dev is not built for this correct. re sandbox infra -> that is not a problem I have. It is more around the ones i mentioned for sandbox simplicity. I would rather have full control over the coding agent setup