Blog
Engineering notes on AI agents, automation, and the infrastructure behind them.
Your Task Manager Is the Best Agent Memory You're Not Using
Agent memory without a new vector DB. Your task app is years of curated, ranked context. ATS gives your agent a hybrid-retrieval channel into it.
AI Agent Model Evaluation: 5 Tests Before the Night Shift
A five-test protocol to catch regressions, compare cost, and canary a new model before it runs an AI agent unattended.
Build the Harness Once With Your Best Model. Run It on a Cheap One.
Agents forget and good ones cost. The fix is not a better model. Put the goal in deterministic scripts and run a cheap model against them.
Most of Your AI Skills Will Rot. Here's Which Ones Compound.
A skill's lifespan is set by what it couples to, not how good the prompt is. Why most AI skills rot, which parts compound, and how to tell.
Claude Code Stops Following Your CLAUDE.md: Read-Once Rules and Hooks
Claude Code reads your CLAUDE.md once at startup, so rules decay as the session fills up. Move the ones that must never break into hooks.
Claude Opus 4.8 Is Out. The Number I Care About Isn't on the Benchmark Chart.
Opus 4.8 shipped May 28. For unattended cron agents, the upgrades that matter are not the benchmark scores. A use-case breakdown from real builds.
Your 50th Skill Makes the First 49 Less Reliable
Past a token-budget threshold, each new skill silently lowers reliability of the rest. Where the work actually lives is below the skill layer.
Self-Hosted Voice AI: Why GDPR Is the Wrong Test (NIS2 Is the Real One)
A GDPR tick isn't a NIS2 test. What you really need to verify with hosted voice AI vendors before NIS2 puts the board on the hook personally.
Splitting Grounding from Reasoning in Browser-Agent Stacks
Browser-agent stacks bundle grounding and reasoning. A local 2B parser splits them, beats GPT-4o on ScreenSpot-v2 by 2.5x, costs $4 to train.
Context Engineering Is Just File Naming
Context engineering sounds new. It is the file-naming hygiene developers always had, load-bearing now because LLMs read what you point them at.
Your AI Workflow Doesn't Need Better Prompts. It Needs Less AI.
Prompting is discovery. Skills are repetition. Gates are how AI workflows become reliable.
Agentic Knowledge Base — Karpathy's LLM wiki, with adapters
A framework that turns whatever task or note app you use into a Karpathy-style LLM wiki. Pluggable adapters, parallel retrieval with RRF.