6 months ago: laptop open all day, typing into Claude Code, waiting. Today: I write a spec, go to the gym, and my phone only pings when an agent needs a decision. 95% of my work runs itself. Full breakdown so you can build your own 👇

Better quality on YouTube: Comment if you want a deep dive on a specific part, or my prompts and skills open sourced.

Chapters: ------------ 0:00 Why I built this 3:13 The tool: agents, environments, skills 6:53 File system & MCP access 9:08 Tasks, templates & the feature pipeline 14:26 The inbox: how agents contact me 16:42 Triggers & automations 19:00 Goals: open-ended loops 21:26 Local runners & costs 23:20 The CLI & YAML config

This is amazing! Thanks for sharing! I effectively built the same thing 😂It's great to see a lot of people converging into the same idea. I find this is the only way to scale autonomy without going crazy 🤪

yeah seeing this a ton lol funny to see everyone converging

@franitog Funnily enough, I had the urge to build something similar but able to get stuff done with a notion database/GitHub issues connected to Claude routines. I interact with the notion ui, the agent pulls via notion connector.

@dannypostma Yeah there are many ways to do it. I thought of using GitHub issues as tasks / kanban but decided to have zero dependencies instead. If you have docker, you can run it. Makes it very portable in enterprise environments

offtopic: so happy that you are back on x

thank you for saying that ❤️

Yessss Danny!! Feel like I am watching your SEO course again, nostalgic haha

ngl was quite fun to record something again!

Finally

sorry i got too wrapped up into my game this week.😂

Watching this as soon as I have a chance! I've got a lot now running with minimal interference, but would be interested on how you do UX/UI. It's the weakest point and doesn't matter how much harnessing, gates, skill files I do, my judgement is still way better than the agent.

using the right skills and agents chained really helps. I just make sure the spec is 100% done by me so that's solid base & i check the plans before implemention

more videos!!

on it! 🤞

I’ve been working on similar set ups, it’s awesome to see what you’ve produced. The real question is: “how do I make it work with the laptop closed” 😂

VM or Claude Managed Agents (expensiveeee)

Well done, but for me this also works with just Codex on Mac, detailed specs + long goal. Meanwhile outdoor i can check progress/ steer Macbook remote through a nice mobile ChatGPT UI, with voice dictation. It also returns videos/ screenshots of new features.

i dont want to steer it

Steering is optional. It only pings when goal is done, or i want to check in with the progress every X hours

Start an education community to get support for what you build. If you ship things like this. It would be valuable and allow others to follow along.

Yeah I’m leaning towards a place to share .md files + videos how to use them.

Try Fusion - it works well for this

He has risen baby

very cool... reminds me a bit of @MCHQ_AI by @pbteja1998 but for coding tasks..

Dude. This is incredible

ty Bud!

DUDE "Sick work bro" ↙ ↘ "Thanks dude" "Legend!" ↙ ↘ BRO ←────────── BRO ↑ "My man" ↓ │ │ "Proud of you" "So good dude" │ │ MOM ───────────→ BRO ↖ ↙ "You guys are amazing" ↺

Hey @yoheinakajima is this like remoko?

I think this is what we all imagined Claude and Codex will be like when we downloaded it. Instead, we got a chatbot with tools

FYI, this only works for indie hackers building throwaway software.

sure Mat

thanks for sharing 🍿

danny is back!

Crazy stuff. Why don't you open source it?

ive been considering it but don't want to maintain or by accident get a security accident due to it

Got it. Looks really cool!

very cool. how does it work with subagents? is this API usage or subscription usage? thanks!

API :cries in high cost: the goal loop now runs on subscription

he's not only back.. he's reinvinted!!

This is EXACTLY what I've been envisioning. Basically a way to work with AI in the same way a large remote team is managed. Any plans to release this? The UI and management layer is so well thought out - and I think that's the hardest part at this point.

Your workflow evolved specs guide agents, and execution only pings for decisions

Similar, but differnet.

95% autonomous still needs that one human decision loop

So you didn't do anything, you literally used a tool, that has stolen thr works from every known source, and you claim victory?

or just get XTRO does all this and more, without wiring everything with duct tape,

What’s the difference between what you’ve built and other systems out there, like Hermes? From what I can tell, I can do pretty much the same thing with Hermes.

Make sure your spec includes clear exit criteria. Agents loop forever on ambiguity. I add a max-step limit and a fallback to human review if confidence drops below a threshold. Saves a lot of debugging later.

Really well made, hoping for a public release

You literally need none of that bs. Claude and codex manage their agents perfectly well. It can run a whole week on a task if needed. The more you add the less it gives you.

what makes it decide to ping you? that boundary is the whole thing. too often and you're babysitting again, too rarely and you come back to three hours of confidently wrong work.