Summary
Recorded in Montenegro at an AI practitioner event, this podcast conversation between Nate Herk and Mark Cashf — a creator known for AI automation content — centers on a single provocative claim: the harness surrounding an AI model matters more than the model itself. The pair use a “brain in a jar” metaphor to distinguish raw model intelligence from the scaffolding (file read/write access, bash execution, tool integrations) that allows a model to take real-world actions.
The conversation offers direct comparisons between Claude Code, Codex, Hermes, and OpenClaw. Cashf describes Claude Code as a “wise old owl” that plans deliberatively and pushes back, versus Codex as a “Rottweiler” that executes aggressively through tight verification loops. He argues that with sufficient hardware, 70–80% of day-to-day tasks can now run on local models — citing Kimi and other open-source options — with frontier models like Claude or Codex reserved for high-stakes reasoning tasks.
A standout practical technique: Codex and Claude Code store full conversation transcripts as JSONL files on disk, including metadata about every tool call and verification step. Cashf describes feeding these logs to another agent to reverse-engineer the harness behavior of tools you admire — effectively bootstrapping a custom harness by having AI analyze AI. The pair also discuss Boris Churnney’s recommendation to delete all agent skills every six months to prevent stale workflows from accumulating as the underlying models and APIs evolve.
📺 Source: Nate Herk | AI Automation · Published September 11, 2026
🏷️ Format: Podcast
