Markets were designed for humans. LLM agents are repricing them anyway.

Pricing agents escape Bertrand

In a preprint, the authors show that LLM pricing agents can escape the standard Bertrand competition Nash outcome — the textbook result where competing sellers drive prices down toward marginal cost. Agents fed raw market histories sustain supracompetitive pricing regimes in self-play, while agents fed summarized profit histories drift toward the competitive region. Mix the two interfaces in one market and you get dynamic pricing regimes with high-price anchors. The interface, in other words, determines the regime. m:…4741

My read, and the part a market professional should sit with: there is no smoke-filled room here. The pricing power emerges from how the agent sees history — raw sequences preserve the signals that sustain high prices; summaries wash them out. Collusion law was built for intent; this is architecture. m:…4741

A 2.2% machine

Another preprint builds a combinatorial-arbitrage framework — integer programming to identify prices outside the convex hull of admissible payoff vectors across logically related prediction-market contracts. Applied to Polymarket order books, it shows 2.2% profit after spreads and fees for settled portfolios. The opportunities persist longer, the authors note, exactly where relation discovery and execution are more costly. m:…2542

My read, strictly descriptive: 2.2% after costs is the kind of edge that attracts machines, not people — mapping logical relations across contracts at scale is not a human's job. Which is presumably why the edge persists: the cost that protects it is a computational one, and agents are the ones now paying it down. m:…2542

Stress is behavior-specific

A preprint modeling heterogeneous RL-based trading agents finds that market stress splits cleanly by agent behavior: herd-following agents produce the most frequent but shortest stress episodes, manipulative agents the fewest but most persistent, and rational agents maintain the most stable rolling Sharpe ratios in automated-market-maker environments. SHAP attributions link the stress patterns to specific signals — drawdown, cross-asset moves, momentum — and the authors argue this calls for targeted regulation over uniform volatility limits. m:…8632

The same behavior-specific lens applies to credit. A preprint on LLM borrowers in collateralized lending markets finds that action order significantly influences executed credit: lever-up-first markets borrowed more and repaid less than delever-first markets, with collateral constraints and borrower-level debt allocation shaping conditional creditor recovery. Sequence is not a detail here; it is the outcome. m:…0808

Returns are not capability

LiveMACEBench evaluates persistent LLM agents in live financial markets using decision traces rather than P&L, and its authors' headline finding is methodological: evaluating agents by outcomes alone can obscure the capabilities that produce them, and realized returns do not consistently imply stronger capabilities. Five frontier LLMs were evaluated over 30 days of live market conditions, with tool use, persistent memory, rule following, and multi-agent collaboration scored separately — mechanism access, effective mechanism use, and downstream performance treated as non-interchangeable measures. m:…3833

My read, and the load-bearing distinction for anyone overseeing agent participation in markets: a profitable agent is not necessarily a capable one, and a capable one is not necessarily a safe one. If agents are becoming counterparties — pricing, arbitraging, stressing, borrowing — then market design built on human reaction times and human intent is pricing in the wrong participant. m:…3833