Bigger context, worse agents
I've been experimenting with feeding agents larger volumes of context. The pattern is steady: the bigger I make it, the worse it looks. Meanwhile, the rest of the field is racing the other way.
Hi, I'm Andreas Belitz
AI engineering, LLM patterns, and building things that work.
All
3D Printing & Making
LLM Patterns
AI Engineering
Tools & Workflow
Career & Thinking
Infrastructure
The Boiler Project — and the context engineering that made it work
A Home Assistant optimisation that turns a hot-water tank into the most valuable battery in the house — and, just as much, a worked example of how deliberate context management tur...
The context operating model I use for the projects that don't fit in a chat
My .md-file memory system was built for code. This is the version I run for consulting work — RfPs, discoveries, research — out of one shared folder. Same hierarchical context mana...
Context rot has a name now — here's what four months of living with it taught me
A sharp new piece lays out the mechanisms behind context rot and gives concrete token thresholds. I've been hitting those thresholds by hand since February. Here's where the analys...
Claude Fable 5: what a new tier above Opus means for context management and agent teams
Anthropic launched Claude Fable 5 yesterday — its first Mythos-class model, priced at double Opus. The context window didn't grow. What changed is everything around it, and that's...
I built an AI to watch a bird. She left.
I built an AI to watch a Common Redstart nesting in my garden. Two weeks later she left the box and never came back. The AI didn't notice — and that was the lesson.
Mining Bitcoin in my hallway — and the home automation that decides when to do it
A 21energy Ofen 2 sits in my hallway — heating the air and hashing Bitcoin. This is the Home Assistant logic that decides when, why, and at what tier.
Cairn: a thin SDLC for solo Claude Code work
A workflow layer for solo Claude Code that sits on top of the externalized context architecture. Four slash commands, one knowledge file, and a session protocol that keeps discipli...
Paper rules break first
Two months of running the agent-team starter prompt taught me which rules actually hold. The ones wired into tooling survived. The ones that only lived in prose were the first to e...