A working notebook on the discipline of harness engineering — the wrapper around the model rather than the model itself. The argument I’m tracking: that the harness defines the productivity ceiling more than the underlying weights. Mitchell Hashimoto crystallised the term in February 2026; within a week OpenAI and Anthropic had published their own treatments, and within two months Martin Fowler’s site had a full-length canonical article on it. The pattern matters: a vocabulary moved from one practitioner’s habit to industry consensus in eight weeks.

Threads to follow:

• Agent = Model + Harness. The simplest formulation, from Hashimoto. Most discussion of “AI productivity” is really discussion of harness quality.
• Context as substrate. The shift from prompt engineering (one-shot wording) to context engineering (the whole information environment the agent operates inside).
• MCP as the neutral protocol. Tools were the bottleneck; MCP made them composable.
• Single agent vs multi-agent. Cognition’s “Don’t Build Multi-Agents” paired with Anthropic’s research-system writeup is the cleanest disagreement in the field — same data, opposite conclusions.
• Evals as steering. Hamel Husain’s argument that without evals you cannot drive the system, only watch it move.