Episode Summary

Dave and Dan break down the harness — the orchestration layer wrapped around an AI coding agent — and why it increasingly matters more than the model underneath it. They get into Anthropic deleting 80% of Claude Code's system prompt with nothing breaking, what that does to your token bill, and the thesis they land on: the better the models get, the smaller the harness you want.

Key Topics

• What a harness actually is — and why Shipwright sits on top of Claude Code instead of replacing it
• General vs. domain-specific harnesses: how Shipwright gets opinionated about PRs and reviews
• Dan's second harness — an upstream, KPI-driven planning layer that writes no code at all
• Boris's reveal: 80% of Claude Code's system prompt deleted, nothing broke, and what that saves you in tokens
• Model economics in practice — Sonnet vs. Opus vs. Fable, and the pull toward Groq and DeepSeek
• Token anxiety: six agents running 24/7, a $2,000 overage month, and what a Fortune 200 team spends vetting open-source models

Notable Quotes

• "Boris said that they deleted 80% of the system prompt for Claude Code and nothing broke."
• "The better the models get, the smaller the harness you want."

About The Velocity Lab

Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field.

Subscribe: RSS