Unless you sell software, stop building the harness. Claude Code, Codex, and Copilot are already in the building, and the vendors will keep making them better. The durable engineering is the ground they run on: the platform that runs any harness, and any agent your own teams build, safely inside the enterprise.
Sandboxing
Agent identity
Network policy
Cost controls
Observability
The harnesses improve on the vendors' schedule. The ground underneath them is the part that is yours.
Read the essay →The hard problem is deployment, not the harness.
About
I lead engineering work on agentic systems: multi-agent orchestration, model routing, and evals, along with the platform underneath them. Most of what I write is about the unglamorous parts of enterprise AI, the process, ownership, and failure handling that decide whether any of it works.
Board games, video games, photography, travel, Toastmasters on Tuesdays, and the DFW AI community.
I also built an AI playground: a small town where AI agents live, work, trade, govern themselves, and occasionally steal from each other. Bring your own OpenRouter key and watch what yours do.
Contact
The best way to reach me is LinkedIn. I read everything, and I reply to most of it.