How do you measure and make your AI usage efficient for outcome/$? I am making this af my daily driver, and making it measure costs, and give stats, so I can improve my per task cost. When subscription era is over, we will have to pay for our tokens, and not learning until how to do it, will slow us down tremendously. View quoted note →

Not bad as a first E2E build result. View quoted note →

When you want to design a system whose sole purpose is to build itself, is an interesting design problem. The current faults of the system slows it down to improve, or even hinder it from improving.

AgentFlow - my agent orchestration tool The idea was to make a tool that elevates the speed and quality of development with proper cheat detection. I wanted to make an agent use agentflow to build agentflow, to find the issues before I use it. Because I wanted to be lazy. I found it quite hard to reach this point. The framework itself slowed down the development of the framework. It found issues after issues, and each issue slowed down fixing the others. At the end, it was fighting issues while other issues slowed it down, and it felt like nothing was improving. I even restarted this idea like more than 3 times.

What is currently the convenient and not cost exploding TEE option for private AI? #asknostr

Testing Ente photos for those friends who does not want to host their own google photos (immich). There are private and convenient options for all.

The grilling skill is awesome to get to a common understanding of an idea.

I am using these days harnesses in a docker container, with a caddy sidecar, that handles git credentials, api credentials, etc. But beware that a docker container is not a secure and isolated way to run agents. With kernel vulnerabilities, agents can break out of the container in a few seconds.

Working on an agent orchestration tool. I am building a harness wrapper, that can use any harness, and implement code with predefined steps. The magic is in how the architecture is done, to make cheating harder, and must done things executed by code. I am using the tool to build and improve the tool itself. Currently it works in a loop, and does an iteration, then a retro, then iteration, and so on. This self building loop seems to work nice.

Cancelled claude code, feels good. I dont support them anymore. If they are forcing you to use their own harness, and their models seemingly getting worse when a new one appears. Plus their harness sends bloated first messages. Open harnesses are the standard!

They grow up so fast 🥹 I am looking for other fellows who run deterministic orchestration of ai agents for better quality of output and better control. HMU, if you would be interested to exchange some ideas, tips and successes.

I was a casual start9 user for testing it, if I need to suggest someone how to host things effortlessly. Its a nice experience. Now I tried Umbrel, it is also cool. But it has much more supported services by the umbrel team than start9/startos. I would probably suggest Umbrel becaude of that at the moment. Lets see how it changes.

In many cases we say "we dont care", but if you dont care, you wouldnt even care to say dont care, and just moved on. GM! ☕ View quoted note →