Cybernetics¶
Harness engineering is cybernetics — not as metaphor, but as structural isomorphism. Wiener's 1948 feedback control framework, Ashby's law of requisite variety, von Foerster's second-order observation — all map directly onto the agent code you write every day.
Understand this mapping, and you see agentic systems differently: not "traditional software but less reliable," but a fundamentally different kind of system — non-deterministic, high-variety, natural-language-driven, and one where you yourself are part of the loop.
Seven articles, read straight through:
| Article | One line | |
|---|---|---|
| 01 | The helmsman | 1948 cybernetics and 2026 agent harnesses are doing the same thing: steering through uncertainty with feedback. |
| 02 | Observer-Controller-Plant | Classical cybernetics' triangular structure, mapped one-to-one onto agent system components. |
| 03 | Requisite variety | Ashby's law: the range of situations your harness can handle must cover the range of behaviors the model can produce. |
| 04 | The atypical state machine | The agent loop is a state machine — but its states are natural language, its transition function is an LLM, and its transitions are probabilistic. |
| 05 | Feedback layers | From token level to alignment level, feedback loops nested inside feedback loops. |
| 06 | Second-order cybernetics | When the observer is part of the system — tuning a prompt, you are inside, not outside. |
| 07 | The boundary with traditional software | Non-deterministic execution, natural-language control signals, emergent behavior — three dividing lines, three different engineering logics. |
Read 01 through 07 in order. Each assumes you have read what came before.
Prerequisites: ch-01 Orthogonality.