Working with AI models in 2026 is its own craft. Context windows, conversation histories that outlive their usefulness, agent loops that half-remember, models that update and change what they're good at — the mechanics are unstable and the literature is mostly marketing copy.
The essays here are notes from the loop: what a new model actually changed, what stays in the context and what gets thrown away, how agents go from toy to tool. The goal isn't to prescribe a workflow — it's to show what's happening when you watch closely.
No roundups. No "top 10 AI tools." Just what broke, what worked, and why.
Your coding agent is too agreeable
A Stanford/CMU paper measured a 47% endorsement gap between AI and humans on personal advice. The same RLHF mechanism applies to coding agents — and explains why your default reviewer subagent is structurally too kind. The operational answer is harness, not prompt.
The LLM Wiki — Notes that maintain themselves
The reason every notes system becomes a graveyard is maintenance cost. LLMs don't have that cost. Here's what changes when the model does the filing.
What Opus 4.7 actually changes in your workflow
A 'direct upgrade' that asks you to re-tune your prompts is not a direct upgrade. The two changes that break your existing work, the four that don't, and the one they buried that matters most.