ohadpr

[ohadpr @ohadpr]

@karpathy There’s a natural generalization of this pattern to include read-it-later lists, ‘subscribing’ to topics without relying on RSS just on an agent to follow and find things of interest to me

[ohadpr @ohadpr]

@bcherny Does it work if you’re already in a worktree?

[ohadpr @ohadpr]

@martin_casado I think there are stronger approaches to secure all sorts of agents e.g https://t.co/nPgV44MntC

[ohadpr @ohadpr]

@sama Is the sign-on bonus open to Clawback?

[ohadpr @ohadpr]

@peduarte Pedro you guys are missing the basics: Let me type something and you figure out if I’m searching for a file or typing an equation or a contact name or whatnot. Having to pick a ‘sub app’ so I can focus Raycast on files/contacts/whatever makes it a no go

[ohadpr @ohadpr]

@gokulr @HarryStebbings The power is in not reading code. In an IDE, it’s unavoidable - files open and pull you in, even with a fully agentic setup.

Drop the IDE and commit to the terminal, and something shifts. Files stay closed, the guilt disappears, and you can focus entirely on driving the agent.

[ohadpr @ohadpr]

@martin_casado @Rafael_L_Spring So git is not required because the code itself is assumed to be generated on the fly, interesting. From prompt to a server running your ephemeral software.

[ohadpr @ohadpr]

@anothercohen Personal tools are the future, I’ve built myself a similar thing combining my most used parts of Obsidian and Soulver

[ohadpr @ohadpr]

Who’s the contrarian saying “Reinforcement Learning good”? 😄

[ohadpr @ohadpr]

@cramforce @thefubhy A human-in-the-loop can be part of a workflow but that doesn’t make the human a workflow.

[ohadpr @ohadpr]

@thefubhy Agents become workflows only in retrospect, once their run is complete. Before they start, Agents have agency, not the rigidity you’d expect from workflows.

[ohadpr @ohadpr]

@trq212 This is a great pattern. Why not dump all tool call results into the filesystem, at the harness level, and let the agent act on the results from there?

[ohadpr @ohadpr]

@trq212 I think devs would find this useful: a sample repo showing a Node server managing multiple Claude agent instances, focusing on keeping them alive and powering multiple front-end chats, each tied to one agent session. Session hydration best practices would be next on my list.

[ohadpr @ohadpr]

Next level!!

[Riverside @RiversidedotFM]

Welcome to the era of chat-based editing.

[ohadpr @ohadpr]

@claudeai Excited to see this concept getting more attention. I wrote about it and built a POC a few months back:

https://t.co/yY69PPKT1A

[ohadpr @ohadpr]

@YaronGalai Very true. In terms of upside imagine OpenAI cracking a Google-like business model

[ohadpr @ohadpr]

@libovness Somewhere around 2k tokens-per-second code and software become obviated https://t.co/yY69PPKlc2

[ohadpr @ohadpr]

@LouisKnightWebb Just heard you on @latentspacepod with @swyx talking @vibekanban. When Claude’s coding and I drift off to browsing, it feels like waiting for code to compile back in the day, same behavior. Funny how that vanished with instant web dev and now came back.

[ohadpr @ohadpr]

I built something with AI.
The model isn't building the app. It is the app.
Voidware. Software that doesn’t exist until you use it.

https://t.co/yY69PPKT1A

[ohadpr @ohadpr]

so cooool

[Riverside @RiversidedotFM]

Introducing the new Conversation Studio.

With full HD live streaming, multi-track editing, AI features, and so much more, it’s definitely a conversation starter.