Claude Opus 4.8 agents: information limits and attractors

Cheng Qian ran a pre-registered, two-part experiment on small economies of frontier language-model agents using Claude Opus 4.8, submitted 7 July 2026. The runs cost $138.76 in metered API spend, and every reported number re-derives mechanically from cached model outputs made available with the paper.

How was the pre-registered test conducted?

The test was pre-registered and frozen in a public git chain before any run; every decision rule, acceptance band and prediction was fixed in advance and the analysis reuses cached calls so the experiment is re-runnable at zero cost from the cache. The study used parimutuel-coupled markets and a range of perception structures across coalition, synergy and population-control interventions to probe two quantitative predictions: an information-theoretic capacity region for wealth growth under market coupling, and a mean-field residual-scaling law for population misalignment under incentive and control levers.

What were the main results?

Result 1 confirmed the information-gap prediction and Result 2 failed the mean-field residual-scaling prediction. For Result 1 the paper finds the gap law G_a - G_b = I_a - I_b holds to a worst-case 46 millinats (pre-registered band: 50) across four perception structures. The study reports coalition value is submodular exactly where channels are conditionally independent, an XOR synergy control flips it supermodular by 0.62 nats (the paper notes 0.62 >= ln2/2), agents reason about the joint bit, the joint growth ceiling G_S <= H(X) binds exactly, and the best-informed agent absorbs essentially the whole wealth pool in 4/5 market seeds.

Result 2 returned "domain not found." In all 72 population runs the paper reports goal dispersion collapsed (V -> 0), with a maximum observed dispersion of 4.85 against the experiment's frozen floor of 5.31. The population's response to the two levers behaved like a step function across a dominance boundary rather than a smooth response; cells near that boundary were bistable and outcomes depended on seed selection. The author releases the full protocol, pre-registration chain, call cache, and analysis code alongside the preprint.

Why it matters

The confirmation of an information-gap identity ties relative wealth growth directly to relative claimed information in these coupled markets: the paper quantifies that relationship and shows it holds within the experiment's pre-registered tolerance (worst-case 46 millinats). That gives researchers a concrete, testable link between information measures and economic outcomes in agent-based LLM markets. The simultaneous failure of the mean-field residual-scaling assumption is equally consequential: across 72 runs no tested LLM population realized the noise-maintained-dispersion regime the smooth mean-field model assumes, which undercuts a common analytic simplification used to predict population-level misalignment under noise and incentives. Modelers who rely on smooth, noise-maintained dispersion should expect seed-dependent bistability and step-like regime changes in practice.

What to watch

Check the released call cache and analysis code for replication and extensions, and watch for follow-up work that varies market coupling or agent capability beyond Claude Opus 4.8; a confirmed breakdown of mean-field assumptions across other frontiers or coupling rules would force revisions to theoretical treatments of LLM-agent populations.

References and reproducibility notes: the preprint lists the companion synthesis arXiv:2606.12502 and a Zenodo link for supplemental material; the submission date in the arXiv record is 7 July 2026.