Posts tagged "handcrafted":

OneLoop: Local First Coding agent written in Rust with nix tooling

I have been working on my own personal coding agent for the last few months. As a matter of fact, my initial commit is on 18th of April 2026 and it went through quite a number of iterations. The git history is quite telling about my experimentation of features such as auto compacting, memory, and multi-model debate/consensus/judgment etc. It supported a few big name model providers on top of OpenRouter initially. After using it for real production work for some time, I am quite happy with the following state of OneLoop [Projects]:

• It defaults to local model (llama.cpp serving Qwen3.6-35B-A3B-Q4_KM.gguf).
• It features 4 tools, on demand skill loading, append only session history with an option to reset via /clear command anytime.
• It has been battle tested, which includes developing OneLoop using OneLoop.
• It supports OpenRouter with web/link tools using /model command

Please note that this coding agent is designed by me for me and I do not recommend you use it for production projects since it lacks any security/permission guardrails. However, I'd recommend you clone the repo and just have fun hacking it. It is super easy to get going if you happen to use Nix on a machine at least with 32GB of ram:

git clone git@github.com:oneness/OneLoop.git
./ols # builds llama.cpp for your CPU/GPU, downloads Qwen model
./ol  # run in another terminal and start interacting with the agent
echo "Give me a mental model how this code base works" | ./ol

I'd like to share a few things I learned along the way building a coding agent tailored to my own workflow. First let me start with things to watch out when you are building your own coding agent, which I highly recommend.

1. Keep your coding agent as dumb/minimal/simple as possible.
2. Prompt engineering requires you to have deep knowledge about the problem domain. So optimize coding agent to teach you using research/learning skills that it can load on demand to instruct agent to help you get clarity.
3. Context management is hard. I am thinking on this. Please reach out if you have novel ideas on how to do this better. I do have a few ideas but not settled yet.
4. When you are in doubt of a feature you want to add, follow point 1.
5. Use local model if possible. It forces you to slow down, be more cognizant of ever-growing context so you can breakdown problems for better solution.

Now, let me talk about how I use OneLoop. Following laws keeps myself and coding agent in check. I retrofitted Asimov's Laws of Robotics [Wikipedia link] as follows:

1. A Coding Agent is not allowed to run in auto pilot.
2. A Coding Agent must be made to obey the SOLID engineering principles.
3. A Coding Agent is allowed to have side effects as long as it does not violate the first and the second law.

Ok, you might say that above is too vague or hand-wavy but trust me, they have served me and OneLoop well. It gave me clarity when I am developing OneLoop in terms of adding or removing features. Here are basic steps I utilize OneLoop for my day-to-day work now.

1. Use OneLoop to get clear understanding of the problem I am trying to solve. This means that I ask OneLoop things like this: "Give me a hello world example of how x y z works". I ask it to generate one off scripts that I can run myself to exercise the code. I change here and there, add print statement and run it to get an intuition of how things work.
2. Then I ask OneLoop to read codes from my repo and give me three options on how to add or integrate that feature.
3. I review all options, ask further questions, ask OneLoop to generate one off standalone scripts so I can run them myself and iterate again from 1.

That is pretty much how I use OneLoop. It supports all I need such as clearing context to start new, summaries of long sessions so I can decide to use or not use in my ongoing conversation, switch to different model if I need (when I am running OneLoop from some low end pocket laptop) etc.

Remember, Coding Agent should serve you, optimize for your understanding, and helping you get better at your craft. If you have been just pressing `Enter` so far, please do yourself a favor and ask your coding agent to ban that command until you confirm that you understand what you are doing. Have fun coding!