Puneet Singh

The best apprenticeship model was always the shop floor. Now the shop floor has an AI teammate. @tobi doing tobi things 👌

Great writeup! Throwing data in a store is not helpful for companies. There is so much more to do to make the data useful.

you can outsource your thinking but you cannot outsource your understanding

Companies like Ramp and Stripe are building software factories using a unified context layer.

@budapp Good domain name 👍

You can text Bud at +1 (415) 525 9024, via Telegram, or use at bud.app. No set up, no downloads, no configuring, no mac mini.

Introducing Bud. The first AI Human Emulator. Bud has a full computer with storage, compute, and memory to build and code, sms and telegram to communicate, a full browser to use, can create/store/edit files, connect and use your tools, learn custom skills, work fully autonomously, and complete any task end to end just like a human. Text the number below or try free at bud [dot] app. Comment for 100k free credits.

Why are all big tech companies building software factories?

@harshilmathur Codex works better for recommendations for deep research as per our testing. You should try it out to compare different models also.

Biggest takeaway: The recommendation is the weakest part, LLMs drift to the safe compromise. Real decisions still need conviction - something LLMs don’t have (yet). But it does show every path. Caveat: full loop on Opus is ~20–30 mins, ~500k–800k tokens.

Built this weekend - autodecision a Claude Code/Cowork plugin that applies @karpathy's LLM Council + autoresearch to business/strategy decisions. Most people ask ChatGPT something, iterate twice twice, and move on. That’s not enough for deep decisions. This: • pulls real data • runs a council of POVs (optimist, pessimist, market, competition, regulator) • models second-order effects • stress-tests worst cases + black swans • finds stable/weak effects → then iterates until something holds up Basically: force the model to argue with itself. Try it on Claude Code/Cowork: github.com/harshilmathur/…

@BadCapitalVC @sundeep I agree that agents + skills + context will evolve to eliminate the legacy UI. However, the agents need a lot of context to understand the data. SaaS is not just the interface or the data, there is a fat business logic layer that always sits above the data layer

Think of the AI stack like a factory: data is the raw material, software is the machinery, and agents are the workers. For decades, owning the machinery was the only moat. But as @sundeep suggested, the "workers" don't need the "machinery" anymore - they can go straight to the "raw materials." It's still early, but it's hard not to wonder what that means for the software layer over time.

@rauchg x.com/xhawkceo/statu… We think “Agents should be portable” and LLM-agnostic.

Whether design belongs in Figma or Claude Design is a distraction from a bigger shift. 1️⃣ Design will become autonomous. More helpful to think of it as 𝙳𝙴𝚂𝙸𝙶𝙽.𝚖𝚍, used by your coding agents running your software factory. 2️⃣ Specialized “personal” design tools generated by teams will proliferate. Design is a capability, not a tool. I agree with @rsms that there are many facets of design, and multiple tools are required. I love prompting in @v0 and it’s become the place where I can channel my inspiration, explore, communicate. But I’m also seeing a new generation of products that use the v0 Platform API or Sandbox and put design on autopilot. There are next-generation agents like @tryflint and trybloom.ai generating design & brand systems and maintaining them autonomously. Flint can even keep your website and content up to date and its design consistent. No human prompting needed. From this we will see the emergence of fully autonomous companies with agents like nanocorp.so and durable.ai, which go a step further and grow and advertise your business. tl:dr; The future looks very different from the present. AI is a true discontinuity. The “here’s the existing thing but with AI and ${jobTitle} is cooked” is short-sighted.

Delightful writeup!

I’d encourage teams to focus on building custom skills on top of LLMs, instead of buying dozens of agents. Here’s why: Most agents are stateless. They rely on prompts and limited context, so every task starts from scratch. Skills change that. Skills are emerging as the way to encode: -domain-specific knowledge -repeatable workflows -company-specific instructions They allow you to teach an LLM how to do something once, and reuse it across tasks, teams, and agents. Instead of managing 100s of disconnected agents, you build a layer of reusable capabilities that compound over time. You share that across the team. In practice: Agents execute Skills provide structure Context makes it all work together The shift isn’t just from tools → agents. It’s from agents → skills, context, and systems. That’s where the real leverage is. We are helping customers build team-wide custom skills captured from deep research on artifacts like source-code, coding sessions, etc.

Your SDLC is already outdated! Part 3 of 3

Your SDLC is already outdated! Part 2 of 3

Your SDLC is already outdated! Part 1 of 3

@rauchg @nicoalbanese10 Software factory is the future! 👍 x.com/xhawkceo/statu…

The software factory is the product: open-agents.dev

@kieranklaassen Compound Engineering Summary 3 of 3. Great article by @kieranklaasse

@kieranklaassen Compound Engineering Summary 2 of 3. Great article by @kieranklaasse2

Compound Engineering Summary 1 of 3. Great article by @kieranklaassen

@ASpittel Wrong question. The future is not interactive. It's event based. x.com/xhawkCEO/statu…

how many coding agents do you feel like you can actually manage at a time?