AI Architecture Updates: July 10, 2026

1. Böckeler Reports Hands-On Results Running Local Models for Agentic Coding

Martin Fowler. In a follow-up to her earlier viability analysis, Thoughtworks’ Birgitta Böckeler documents hands-on experiments running local LLMs for agentic coding on M3 Max (48GB) and M5 Pro (64GB) machines, testing Qwen 3.6 35B MoE, Qwen Coder Next 80B MoE, and Gemma 4 variants, and landing on a 4-bit quantized Qwen 3.6 35B as a workable but RAM-tight production configuration. She reports high variability and confusing results, including identical configurations yielding sharply different success rates across machines and failure modes such as memory crashes in long conversations, weak handling of large single-file context, and incomplete implementation of complex requirements. The practical conclusion is architectural: local agentic models fit small, pre-planned tasks best, benefit from explicit file-pointing over code search and fewer edits, and still make human code review essential. Source