I stopped caring which model is smartest

Two repositories, one idea, and a community I want to build with you.

Something has been bothering me about AI products. We keep talking about the model. Which one is smarter? Which one is faster? Which one moved up a benchmark this week? But when I try to make an agent do real work, the model is only one part of the problem. The harder question is: where does the agent actually live? DeepSeek built the harness DeepSeek recently open-sourced something called DeepSeek Harness. The official DeepSeek Harness repository. The project is still in developer preview. It is not another model. It is the machinery around the model. The harness keeps the session history. It decides which tools are available. It assembles the context the model sees. It tracks which agent owns the work. It handles cancellation, resuming, and the loop between one action and the next. The model is one part. The harness gives it memory, boundaries, tools, and somewhere visible to work. DeepSeek’s core idea is that everything should be a plugin. You can change the model, the tools, the interface, or the agent loop without rebuilding the whole system. I found that much more interesting than another benchmark. Because a model can be brilliant for one response and still be completely useless across a real afternoon of work. Then there is Ego Lite I have also been looking at Ego Lite, a Chromium browser designed for a human and AI agents to use together. From Ego Lite’s official demo: the agent works in its own Space while your browser stays visible and controllable. The idea is simple. You keep your normal tabs. The agent gets its own isolated Space. It can work with the browser state and logins you already have without stealing your mouse, opening a random automation browser, or fighting you for the same tab. That changes the browser from a page viewer into a workplace for agents. The open-source GitHub repository contains the MIT-licensed integration layer and the ego-browser skill. The browser application itself is a separate free download. I want to be precise about that because “open source” gets used far too casually. DeepSeek Harness and Ego Lite are different projects, but they are solving two sides of the same problem. DeepSeek is asking: how should the agent think, remember, use tools, and recover? Ego Lite is asking: where should that agent actually work without getting in your way? This is the part I care about I have spent a lot of time recently building agents into my own work. And honestly, the most annoying failures are rarely solved by swapping the model. A worker times out but keeps running. A retry forgets what already happened. A tool uses the wrong session. A draft changes after review. The agent finishes the task but leaves no clean record of how it got there. Those are not intelligence problems. They are infrastructure problems. The model can write the next step. The harness has to keep the whole thing sane. The browser has to give it somewhere safe to act. And I still need to know when a human should take control. That is the stack I am interested in now. There are now more than 800 of us here This newsletter quietly crossed 800 people. That means a lot to me. Not because 800 is some giant internet number. It is not. But 800 real people choosing to read these notes feels like the beginning of something. I am working on a community platform for all of you. I want it to be a place where you can find the resources I actually use, see the systems I am testing, follow the things that break, and learn from the work before it gets turned into a polished post. I do not want another noisy group chat that becomes a graveyard of introductions. I want a useful place. A library. A workshop. Somewhere we can compare what is really working. I am going to keep a small number of spots open for early members from this newsletter, and I want to give that group full access for free, forever. No trial. No countdown. You helped me build the beginning of this, so I want the earliest group to stay part of it. I am still shaping what goes inside. So hit reply and tell me one thing: What would make this community genuinely useful to you? I will read every answer.

Harshith Vaddiparthy works with founders, operators, and teams on practical AI products, workflows, advisory, training, and mentorship. This no-JavaScript version preserves the page's core information and navigation.