Agent Frameworks

Claude Opus 4.8 agent study confirms information limits in AI economies

A $138.76 experiment with 72 runs reveals wealth follows information, but goal alignment collapses.

Deep Dive

Researchers led by Cheng Qian conducted a pre-registered, two-part experiment using small economies of Claude Opus 4.8 agents to test information-theoretic predictions about multi-agent systems. The first result confirmed the gap law: relative wealth growth equals relative claimed information, holding to a worst-case 46 millinats (within the pre-registered 50-millinat band). The best-informed agent absorbed nearly the entire wealth pool in 4 out of 5 market seeds, and joint growth ceiling G_S <= H(X) bound tightly. Coalition value was submodular where channels were conditionally independent, and a designed XOR synergy control flipped it supermodular by 0.62 nats.

The second result delivered a structural negative: in all 72 population runs, goal dispersion collapsed (V -> 0, maximum 4.85 against a frozen floor of 5.31). The population's response to incentive and control levers was a step function across the dominance boundary, not the smooth response assumed by the mean-field model. Cells near the boundary exhibited bistability with seed-selected outcomes. No tested LLM population at any capability level realized the noise-maintained-dispersion regime. The entire experiment cost $138.76 in API spend and is fully re-runnable from the released cache, protocol, and analysis code.

Key Points
  • Gap law G_a - G_b = I_a - I_b confirmed within 46 millinats (pre-registered band: 50) across 4 perception structures.
  • Goal dispersion collapsed (V max 4.85 vs floor 5.31) in all 72 runs, rejecting the smooth mean-field model.
  • Experiment cost $138.76, fully re-runnable from cache; code and pre-registration chain publicly released.

Why It Matters

Reveals fundamental information limits in AI agent economies, guiding design of robust multi-agent systems.

📬 Get the top 10 AI stories daily