Jay Taylor

@akkkkiira Isn't that a good thing? :D

that feeling when a potential client you were eyeing comes to u and wants to use ur product really bad 😩

@paulpgustafson @karpathy @maxbittker Try out github.com/jaytaylor/typi… > typing-bird watches a tmux pane for idle output, then sends the next message (plus Enter) to that pane. It cycles messages forever and can inject itself as a bottom 5-line pane in a target tmux session.

@karpathy @maxbittker Did you figure out how to submit prompts (“press enter”) to Claude code programmatically through tmux? They blocked all the straightforward ways to do this

From @karpathy's autoresearch .md

@akkkkiira @ArgaLabs Way to go Akkira, I dig it :)

Introducing @ArgaLabs' newest feature: red-team your agent, hosted on any software, autonomously. We built infrastructure that enables us to recreate a behavioural clone of any app, aka Digital Twin (inspired by @jtaylor), extremely fast. Here’s an example of Arga testing newly-added guardrails to a bot in our Discord digital twin. The clones have full official API/SDK support, are fully functional, and you can directly preview your changes on the Arga website. We also support red-teaming on your app — where we spawn a swarm of agents into your app, hosted in our remote sandboxes. Our agents understands your program by gathering external context (we support almost every mainstream platform out there) to generate user stories, and they click through like real users to uncover unexpected behaviour. Let our agents find bugs, because you were never meant to be QA for AI.

@xAviation Fascinating, more reading: nationalinterest.org/blog/buzz/why-…

The U-2 always quoted as one of the hardest planes to land.

@simonxabris @Cloudflare Yes - you've discovered factory.strongdm.ai/techniques/fil…

Figured out how to do perfect agentic search for all transcripts of a user in kopia Currently the agent does classic RAG. The way better approach is to mount the transcript files into @Cloudflare sandbox and give the agent a tool where it can run shell commands in that agent.

@Stand_Frame @official_taches What is your workflow? With `Sprint Plan -> Execute -> Refine` meatspace can scale to 4-8 useful tasks in parallel

@official_taches What the fuck dude! I have a VC who says he’ll only speculate if I’m burning a billion tokens per day. 70% reduction totally blows that. What am I going to do now?? I already showed him three TUIs going at once, more monitors I guess, more directories 😉

I'm having one of those kind of breakthrough days that led to the original GSD. I've just HUGELY improved the quality of the standalone GSD system, whilst making it much, much faster and much, much cheaper to run. Token usage is down ~70%! None of this would have been possible if I kept developing GSD as an "add-on" to tools like Claude Code. Fuuuuuuuuuuuuuuuuuuck.

Even OpenAI is adopting NL Spec as a distribution mechanism: > Implement Symphony according to the following spec: > github.com/openai/symphon… The origin of the Natural Language Spec concept is github.com/strongdm/attra…

@Stand_Frame @strongdm @philhassey What am I looking at? :)

@strongdm @philhassey 🤙

@JamieOarton Thanks for the shout out, Jamie! Feel free to drop me a line if you have any questions.

StrongDM, the cyber security software, only have THREE devs now as they re-engineered their workflows to optimise for AI. VERY few others are operating at this level yet… #ai #chatgpt #code #dev

@akkkkiira I tried it out and like the direction! Have you considered keeping the editor and adding an interactive agentic dotfile builder as the initial page?

Just built a tool to create workflows for agents without dealing with orchestration. It’s a browser-based visual builder for making workflows across coding and non-coding agents. It implements StrongDM’s Attractor framework, and helps you build agent harnesses easily. We currently use it internally for CRM data analysis at @ArgaLabs but curious to hear about any other use cases. u can check it out here: github.com/ArgaLabs/attra…

@simonxabris Really cool work, what is going on with the file browser? Independent totally new component?

ghostty++ is evolving. Add a terminal overview with keyboard shortcuts so I can easily navigate between terminals in different projects.

@spring_stream @simonw Cloning existing SaaS is straightforward when you have a portfolio of high-quality client SDKs on hand. For greenfield, thorough planning and auto-analysis+issue remediation via e.g. Attractor is becoming increasingly unreasonably effective. #community" target="_blank" rel="nofollow noopener">factory.strongdm.ai/products/attra…

Do you have any thoughts about how they deal with insufficient precision and ambiguity of the specs then? I’m not even talking about non-functional aspects or implementation choices, but the behavior and UX of the thing being built. If it is a workflow-like product or a product from an established category - fine, everything is known. In cases of “something new which hasn’t been built before” - do they let AI fill the gaps and make choices? or do they (humans) refine and expand the set of scenarios every time AI declares it’s “done” and it turns out to be not the thing they wanted? I like the spirit of what they are doing but still looking for a functional distinction from “project manager throwing the waterfall spec over the wall to a team of SWEs and QAs”.

I wrote about the most ambitious form of AI-assisted software development I've seen yet - Strong DM's "Software Factory" approach, where two of the guiding principles are "Code must not be written by humans" and "Code must not be reviewed by humans" simonwillison.net/2026/Feb/7/sof…

@twistartups @mvanhorn thanks to @mvanhorn, we enjoyed watching your piece here at the StrongDM-AI HQ!

LIGHTS OUT! That is the future of software factories says @MVanHorn. The rules: no code to be written by humans, no code to be reviewed by humans. PERIOD! How long before we have software companies with human OUT of the loop?

@brunotorious @strongdm @BuiltByJustin Very cool Bruno, way to go!

I ported @strongdm 's attractor to ruby because I could... github.com/bborn/attracto… cc @BuiltByJustin

Send "Someone told me the filesystem is all we need and that I can solve problems and make systems reliable by throwing more tokens at them. How can we make this more agentic and implement pyramid summaries as well? Look at the tenets here: factory.strongdm.ai/llms.txt" to Claude

If you are among the first to get it up and running please drop me a line!

Thanks for sharing my Digital Twin Universe project!

Still waiting to see who gets it up and running first!

@peteskomoroch @simonw As the person who created this I can confirm you've got the right idea. The Digital Twin Universe continues to be extremely useful and pay dividends for developing with agentic coders at scale.

This reminds me of a dev strategy I like: building prototype apps in a downsampled version of the database where the universe is limited to data from a subset of all users. I also mock important 3rd party APIs, which makes debugging much easier, but StrongDM is doing this for everything with LLMs:

@theo Boom

GPT-5.3-Codex and Opus 4.6 both just dropped within a few minutes of each other