Chapter 8: Why Infrastructure > Intelligence
The billion-dollar assumption: pour money into model training and agents get better automatically. But the data tells a different story. The harness matters more than the model inside it.
Same model. Same week. Two different harnesses. GPT-5.5 in Codex scored 61.5% on SWE-bench. GPT-5.5 in Cursor scored 87.2%. That's a 25-point swing from the wrapper alone.
A model upgrade typically gains ~8 points. The harness gained 25.
A thin harness: model + basic tools + simple loop. 3 components. A thick harness: context management, error recovery, tool design, permissions, memory, observability, guardrails. 9 components. The thick harness wins benchmarks by a landslide.
A GPT wrapper is barely a harness, just API calls. ChatGPT adds conversation management. Claude Code adds agentic tools, context management, and permissions. Cursor goes furthest with deep IDE integration, custom indexing, and multi-model orchestration.
Not all harness components matter equally. Click each bar to see why it matters. Context management and tool design alone account for 60% of the impact. Get those right first.