OneLoop is a personal, local-first coding agent built in Rust using Nix tooling, running llama.cpp with the Qwen3.6-35B-A3B-Q4 model. The author shares lessons learned building it: keep agents minimal, use local models to slow down and manage context, and avoid autopilot usage. The workflow involves using the agent to understand problems, explore options, and iterate with standalone scripts rather than blindly accepting generated code. The project is open source and easy to run on Nix with 32GB RAM, though it lacks security guardrails and is intended for personal hacking rather than production use.
Questions this post answers
What model does OneLoop use by default for local inference?
OneLoop defaults to llama.cpp serving Qwen3.6-35B-A3B-Q4_K_M.gguf locally. It requires a machine with at least 32GB of RAM. It also supports OpenRouter with web/link tools via the /model command for switching to cloud-hosted models when needed, such as on lower-end hardware. Developers building local AI coding setups track hardware requirements and model choices on daily.dev.
What are the key lessons learned from building a custom coding agent?
Keep the agent as minimal as possible, use local models to force slower and more deliberate context management, and invest in prompt engineering grounded in deep domain knowledge. Context management is the hardest problem. When unsure about adding a feature, default back to keeping things simple. Local models also help break problems into smaller, clearer pieces. Engineers designing their own agent workflows share patterns and pitfalls on daily.dev.