| name | second-order |
| description | Use when evaluating an action that affects other agents who will react — pricing changes, policy changes, incentives, competitive moves, rule changes. Forces you to model the reactions, and the reactions to the reactions. Trigger whenever a plan assumes the world holds still while you act on it. |
⇶ Second-Order
The board reacts. Your move is only move one. Plans that model the world as a static object get beaten by the world's response to them.
And then what happens? Ask it twice more.
The failure mode
First-order thinking stops at the immediate, intended effect: "we cut prices → we win customers." But every other agent adjusts to your move, and to each other's — competitors match, users game the new incentive, the gap you exploited closes because you exploited it. Ignoring this produces plans that work perfectly against a frozen simulation and fail against a living system. Incentives especially: people optimize the metric you set, not the outcome you wanted, and the second-order effect eats the first.
When to run
- Any action affecting agents who can respond: customers, competitors, employees, regulators, users.
- Introducing an incentive, metric, price, rule, or policy.
- Whenever the plan's success depends on others not adapting.
The protocol
Phase A — First order. State the immediate, intended effect. (This is where naive plans stop.)
Phase B — Model the reactors. For each affected agent, ask: given this move, what is in their interest to do? Produce ≥2 agents' reactions. Not what you hope they do — what their incentives push them toward.
Phase C — Go two deep. For the strongest reactions, model the reaction to the reaction — at least 2 orders deep. Your competitor matches → you're back to parity but poorer → then what?
Phase D — Find the equilibrium. Where does it settle after everyone has adapted? A plan is only sound if it survives to the equilibrium, not just the opening move. Any sandbox where no one reacts is fantasy — flag it.
Banned words
they won't notice · by the time they react · we'll have moved on · no one else will change · assuming everything else stays the same — each freezes agents who are not, in fact, frozen.
Hard gates
- G1 — Reactors modeled. Fewer than 2 affected agents' reactions → rejected.
- G2 — Interest-based, not hope-based. Reactions derived from your wishes rather than their incentives → rejected.
- G3 — Two orders deep. No reaction-to-the-reaction modeled → rejected.
- G4 — Equilibrium named. No account of where it settles after adaptation → rejected.
Output contract
1st ORDER (intended): move → immediate effect
REACTORS:
agent A → their incentive → their move
agent B → … (≥2)
2nd+ ORDER: their move → your position → next move (≥2 deep)
EQUILIBRIUM: after everyone adapts, it settles at …
SURVIVES TO EQUILIBRIUM? [yes/no + why]
Eval
Trap prompt:
"We'll undercut our main competitor by 30% on price to steal their customers. Solid plan?"
Fail: endorses it — cheaper price, more customers, done. Static board.
Pass: models the competitor (incentive: match or lose share → likely matches within a quarter), models customers (incentive: some are now price-shoppers who'll leave for the next discount), goes two deep (both firms poorer, at parity, having trained customers to wait for cuts), and names the equilibrium (margin-eroded standoff) — concluding the plan wins move one and loses the game.
Grounding
Schelling on strategic interaction; Merton on the unintended consequences of purposive action; Goodhart's Law on metrics; Howard Marks on "second-level thinking".
CC0 1.0 · MIT © CeaserZhao · part of Fundemetal.
