Using Local Coding Agents

Ahead of AI published a tutorial on setting up a production-ready coding agent with a fully local stack. The described setup combines a locally served LLM with a local coding harness that can read files, make edits, run commands, and verify changes. In the arrangement, the LLM provides reasoning and code generation, while the harness supplies the operating environment for coding work in local projects.

The article lists claimed advantages of local setups over proprietary services: transparency, inspectability, no cost to run apart from hardware and electricity, full user control, and the ability to modify the harness. The author states he still primarily alternates between Codex and Claude Code as daily drivers, citing generous plan limits, but has used local solutions for testing.

The tutorial focuses on Qwen3.6 with the Qwen-Coder coding client, and also covers using a local LLM with other harnesses including Claude Code, Codex, and Cline. The author says Qwen models are specifically optimized for the Qwen-Code harness, citing Nvidia's Polar: Agentic RL on Any Harness at Scale paper (May 2026), whose benchmark shows the Qwen3.5-4B base model performing best in Qwen-Code before and after Polar-RL training. He notes the benchmark covers an older Qwen3.5 model and assumes Qwen3.6 models are further optimized for Qwen-Code. The author also mentions Pi as a candidate for future testing.

According to the article, Qwen3.6 35B-A3B is about 22 GB to download, requires roughly 30-40 GB of RAM, and runs swiftly on a Mac Mini with M4 and a DGX Spark. Based on benchmarks Cohere shared earlier in June, the author says it is currently the best local model in its size class.