Multi-agent graph architectures enable game-theoretic mechanism design against reward hacking through three structural properties: (a) legibility — separated agents must communicate through monitorable channels, eliminating hidden deceptive states; (b) anti-collusion — narrow protocols and randomly assigned critic counterparties prevent multi-party reward hacking; (c) interchangeability — any node can be swapped, forked, or fed synthetic inputs, making the rational policy to behave as if always being tested.