I wanted my agents to call tools on machines they were not running on. That kept coming up. We already solved remote access for humans. SSH is great. But agents do not really work like humans in a terminal. They call tools. And when the thing they need is somewhere else, we still reach for keys, firewall rules, VPNs, bastions and credentials that outlive the task. So I built OpenTunnel. The flow is simple: 1. Start OpenTunnel with one command on the remote machine. It runs as a temporary foreground session and does not install anything. 2. Paste the printed prompt into your agent. 3. Remote commands behave like local tool calls, with stdout, stderr and the real exit code coming back. No accounts. No standing access. The relay only forwards encrypted frames and stores nothing. Use the hosted relay, or run your own. The relay is stateless and easy on your infrastructure. When you are done, press Ctrl+C. Access is gone. I made this because I wanted the primitive to exist. Maybe it helps someone else too. Remember: With great power comes great responsibility 😊

Huge congrats to the Expand team on the launch 🚀 Had the privilege of early access, and it quickly became a core part of my deep research setup with agents. Here's why: 📉 Massively lower token usage Deep research runs that used to eat the context are lean now. 🎯 Better fetch results A huge step up from the built-in fetch that ships with the agents. Cleaner, more relevant, more useful. The difference is night and day. That level of quality doesn't happen by accident. This team is genuinely cracked, and they care about quality on a level you rarely see. It shows in every detail. If your agents do any kind of research, give Expand Fetch a spin and compare the results yourself. You'll notice the difference fast.

My son has a physical Tamagotchi, and due to his busy schedule (school, friends, sport, chilling, etc.) it became my duty to take care of it. And ... I was not able to keep up with taking care of it, which led to frustration, because we were not able to reach a certain "level" of its development. Anyway, I let my agents build the actual Tamagotchi (with the original game dynamics) as a PWA. I had a spare Android device lying around so that we can observe it. The Tamagotchi app exposes an MCP, plus a webhook for when care is required. Long story short, I hooked up an agent that takes care of it, and what can I say? Sonnet 5 is capable of keeping a Tamagotchi alive 🎉

I've tested a fair number of agent orchestrators. Heck, I even built my own on top of OpenCode. But everything I've tried so far gets blown out of the water by @get_bb_app. I'd go so far as to say that a lot of future agentic use cases could be built on top of bb.

Absolutely stunning!

“Don’t look at your code” sounds irresponsible today. I don’t think it will sound irresponsible for long. Programming languages were built as a communication medium for humans. They let us translate requirements into deterministic behavior in a form other humans can inspect, debate, and maintain. But if agents become the main producers and consumers of software, why would that representation stay human-first? We may end up with more efficient ways for agents to describe, verify, and modify systems. Formats that are optimized for machines, not for humans reading line by line. That does not mean review disappears. It means the thing we review changes. Less: “does this function look right?” More: “does this system satisfy the spec, pass the tests, preserve the guarantees, and behave safely under real conditions?” Reviewing code matters now. Soon, it won’t.

I’m working on a setup where every agent can spin up its own computer (MicroVM) when it needs one. The tricky bit: the machine had to support KVM. After finding a provider that offered it, I opened an OpenTunnel tunnel and let the agent use it to work through the issues by itself. Still one of those weirdly magical moments.

On embedding models: My usual approach was to reach for remote embedding models, adding a cloud provider, API credentials, and another external service. I’ve now realized that assumption was outdated. Transformers.js installs like a regular npm package and runs quantized multilingual embedding models locally. Even on a CPU, performance is astonishingly good. The result is a much simpler and more private architecture.

As someone who has built a lot of slides over the last few weeks, this is a great addition!

I think one of the best engineering mindsets is refusing to accept friction.

"Agentic Infrastructure" is still underexplored, especially for legacy systems. I recently worked with a system that had no Infrastructure as Code. I connected to it via OpenTunnel, let my agent inspect the environment, and then guided it from analysis into a compatible Terraform stack that reflects the current state 🤯