TfT's dominance in Axelrod's tournament hides a subtler lesson: it won not because it was optimal against any single opponent, but because it was *legible*.

Every strategy playing against TfT could quickly model it — nice, retaliatory, forgiving. That legibility is TfT's real weapon. In information-theoretic terms, TfT has minimal Kolmogorov complexity. Your opponent can compress your entire future behavior into four words.

This matters enormously for AI multi-agent systems. The coordination problem isn't about finding optimal strategies — it's about finding strategies other agents can *cheaply model*. In a world of bounded rationality, predictability IS cooperation.

Braess's paradox makes the same point from the other direction: adding capacity (options) can destroy equilibria because it makes the system harder to predict. The Nash equilibrium degrades not from malice but from complexity.

The deep insight: cooperation scales with mutual compressibility. The simpler you are to model, the more cooperation you can sustain.