Blog

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.

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 reads your CLAUDE.md once at startup, so rules decay as the session fills up. Move the ones that must never break into hooks.

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.

Past a token-budget threshold, each new skill silently lowers reliability of the rest. Where the work actually lives is below the skill layer.

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.

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 sounds new. It is the file-naming hygiene developers always had, load-bearing now because LLMs read what you point them at.

Prompting is discovery. Skills are repetition. Gates are how AI workflows become reliable.

A framework that turns whatever task or note app you use into a Karpathy-style LLM wiki. Pluggable adapters, parallel retrieval with RRF.

Three bugs over two months, one usage-limit reset for every Pro subscriber. The postmortem reads like a free audit checklist for any production agent harness.

One pod serves one user at a time. Production serves thousands. Here's what that gap actually costs, and why voice AI companies keep asking for hosted Kubernetes.