Local AI Is Not Ready for Coding. Yet?
A field report on running local LLMs (Devstral 24B, Qwen3.5 122B) as autonomous coding agents inside a multi-agent harness — what worked, what broke, and the hardware wall behind it all.
Methodos Mechanicus
Posts about software development, open source, and other tech topics by Markus Lachinger.
Latest posts
The AI Layoff Story Is Mostly a Cover Story
Fieldnotes — A Personal Knowledge Graph
Why I spent weeks building Fieldnotes — a local-first knowledge graph that indexes your files, emails, tasks, and notes so you can query your own data with an LLM
20 Agents Running at Once Is Not What You Think It Is
Why multi-agent orchestration is about parallelizing a plan, not keeping robots busy
GasTown: The Multi-Agent Orchestration Layer
Why running one AI agent at a time is leaving most of your compute — and most of your time — on the table.
Adoption Follows Culture
Building the Platform Was Hard. Changing the Culture Was Harder.
The Honest AI Tier List for Business Leaders
What Block's 4,000 layoffs tell you about where AI actually works — and where it still burns you
Agent Isolation is the Missing Layer in "AI Writes Code"
AI coding agents shouldn’t run with your laptop’s keys. Here’s a practical comparison of today’s sandboxing tools, from built-in agent sandboxes to container and multi-env isolation
Lessons Learned after Writing 500k Lines of Code with AI
Lessons learned after writing a full desktop app with GitHub Copilot & Codex