Code w/ Claude & Claude Founders House – Review

Transkript

Reviewing the Code w/ Claude keynote

The pattern behind the most successful startups and companies is that they build for something that is still not here. They look at what is possible, build for it, and try to follow the exponential.

It was a nice touch that Boris opened with a family story: back in the Soviet Union his grandfather, one of the first programmers there, worked with punch cards, and his mother used to color all over the stacks of cards he brought home. Boris himself first learned to program on a scientific calculator, mainly to get better scores on his math tests.

His core point was that for about seventy years programming meant writing software by hand, and that in the last year or two it changed fundamentally: we went from source code to managing agents, and now we are moving again to managing loops and routines. The agents keep getting more independent and can increasingly work autonomously on their own for longer stretches.

On the question of what is next and what is not here yet, my impression was that Boris suggested the context window will essentially become unlimited. I do not know how they will do it, but that was my takeaway. And secondly, that memory will be solved, at least from a user's perspective. The context window in particular is really interesting.

What is new in the Claude desktop app

Lately I had a billing bug where usage kept going to the API, and after a short period I had burned about 300 euros and had no tokens left. So I was more or less forced to use Claude in the desktop app, and it turns out it is pretty great.

You do not need several terminal windows anymore: you can stack your sessions in the sidebar. You can see the screenshots that Claude takes right inside your window, and you can comment on code or on the CLAUDE.md file and let Claude explain things or just change them.

The second thing I found really useful is easier terminal handling. People often struggle when they need to delete or copy and paste something in the terminal. If you switch to fullscreen rendering, you can actually use the mouse in the terminal and copy and paste with it. The magic words are the /tui fullscreen command, and /tui default switches back. You get mouse support (click and drag to select, and it copies automatically), and the documentation lives at https://code.claude.com/docs/en/fullscreen for the full details.

The desktop app essentially works like an IDE, similar to VS Code. You can open a terminal, open files, and edit them directly. Sessions that would normally live in a terminal window can just be activated; you add a comment, tell Claude what to do, and it takes that into the prompt. There is even a preview, so you can flip through files and edit them directly, which is much more comfortable than a bare terminal.

Remote sandboxes, managed agents and memory

Remote sandboxes were discussed a lot and are especially interesting for enterprise customers. You can run Claude on your own infrastructure, or in a container on Anthropic's infrastructure with your model there, inside your enterprise, and you can have secured, authorized MCP servers behind your firewall.

Managed agents are fairly straightforward, but some people have a hard time finding them, because they are not in the app: they are in the Console. With templates you can generate agents relatively easily. It works like asking the user questions: Claude prompts you for your intent, and then it builds a small agent.

You also do not need to keep your secrets in a separate secret manager that is hard for the agent to reach. There is a vault where you can store secrets, and the agent has memory. The newest part is that, like us humans sorting through our memories while we sleep, the agent can sort through and clean up its own memory. That is the "dreaming" step.

There was a nice example from the workshop that I wanted to show: a research desk that works as a deal desk for analyzing M&A targets, with some example companies to work from. It is open source in Anthropic's Code w/ Claude workshops at https://github.com/anthropics/cwc-workshops/tree/main/research-desk and makes a nice blueprint for your own agents.