Context Engineering vs Harness Engineering vs Software Engineering

The genuinely new concept in harness engineering is reinforcement learning a model specifically for a particular harness's tool interface (8:45). For example, Codex's apply-patch format and Claude Code's edit tool are incompatible — a model RL'd for one cannot use the other. This tight coupling between model weights and harness tools creates a competitive moat that pure context engineering cannot replicate.

The panel converges on the idea that every layer of agent intelligence — from basic tool-calling agents to sub-agents to orchestrators like Gas Town — is fundamentally just another while loop wrapped around the previous one (12:30). Each additional loop adds a layer of autonomous work capacity, and the decision of when to add another layer versus optimizing the current one is the key engineering judgment call.

Vib draws a powerful analogy between hand-writing assembly code versus trusting the compiler and hand-tuning prompts versus trusting frontier lab teams (17:30). Most engineers cannot beat the compiler (Anthropic/OpenAI's 40-50 person optimization teams), but occasionally domain-specific knowledge about data patterns or cache locality creates genuine alpha worth pursuing.

The panel challenges the traditional bitter lesson (that engineering around models is wasted effort) by arguing that code is now cheap to produce and delete (22:00). The real skill is "surfing the models" — continuously applying context engineering techniques that yield 5-10% improvement over naive use, staying ahead of each model release cycle rather than being made obsolete by it.