Try PostHog yourself: https://go.posthog.com/david

Cloudroom: https://cloudroom.dev/

so funny hearing normies saying AI is dogshit while they're using the free version of chatGPT 😂

"remove ourselves from the loop". and so it begins... :face-orange-biting-nails:

im glad im not the only one whos having multiple agents being run like how you show at 16:20 although i think ive figured a way around that but it relies on running everything in a linux subsystem and setting up what i like to call "characters", at least thats what im building right now and the characters being built saves me endless amounts of time, the linux subsystem is taking a while to be built, like its probably the longest project ive ever had and once i solve this headache it will give me more time to spend playing video games instead of having to constantly monitor my agents XD

always great videos!!

Looking forward to trying it out, nice work.

Am working daily with ai since the github copilot preview. Agents are worse in every work than humans, they are just faster

you can create an agent observer of your actions outside of harness

Paradox with all those factories and long running stuff is that on intenet there is no resources (of which im aware of) about what survives after compaction, what and to what extent survives, how reliable is the compacted orchestrator and actually those are the questions for any prolonged orchestrator task for coding sessions running weeks. Even the skill loading logic which depends on harness and target llm model(!). I sometimes doubt the guys who run those things do really have a look into what runs for prolonged times. Context7 solves only some of the other issues. What you do when llms write plans and code out of memory which is stalled and code gets written with outdated api or function calls which will result in runtime errors say for TS or python. Do YOU make sure subagents dont silently change the specs ( I saw!) in order the inplementation to comply with changed specs. I saw tests being written AFTER implementation was done, while being instructed to do TDD. There is so much devil in the details. And all I see some wonder factories and stuff.

i was developing a software factory about 6 months ago. the issue is the self healing. a.i. will just not stop. so you end up burning tokens on bullshit.  im still having it fix stuff that it built in from then.

Hopefully the big companies take a break and then start again. Could use a small company to start out right, switch to fattening up for a little while, then switch to a small friendlier company after they got their feet wet. Could be a cycle or many, before events are established. 💯:goat-turquoise-white-horns:🗺

Glad i started cheap lol.. Thanks for sharing!

Hi David, great vid, great platform idea. I'd be down to talk shop regarding your proposa. I've been doing nothing but multi-agent setups since ~Nov 2025 - completely obsessed with it. And currently looking for a feasible way to power my swarm on cloud infra. Still running on Hetzner with mix of trad VPS and k3s atm, but feels "anachronistic" and bloated. As of late looking more into CFs isolates thing - especially since Theo's TSRust thing happened and CFs move to buy Demo and incorporate celld. So would love to chat 🙂🤟🏻

2023: AI just makes stuff up!
2026: AI can't make stuff up!

30:16 Devin now allows login with Chat GPT OAuth

Ultimate evolution of Vibe coder

um'... yah' got somthing stuck on your nose, right about the middle. Yah might wanna wipe that off??

You're describing GitLab at 28:22

The bigger question is, when will I be able to upload my consciousness to the collective / hivemind?

The transition from non-AI coding (just one of many other fields of AI control) to AI agent wranglers will be transferring all of the power of software development from the many software companies into the hands of the few AI cloud providers. Centralization is very bad. Have you seen what governments and corporations have done with such centralized powers?