A research agenda

1. 1

   Common estimands would separate configuration effects from instrument effects.

   A shared financial-task battery should compute the same disposition, anchoring, extrapolation, loss aversion, and probability-weighting statistics using multiple instruments, payoff-equivalent wordings, and dated releases. Wording sensitivity and model drift imply distributions rather than single calibration values (Dong et al., 2026; Gao et al., 2025; Wang et al., 2025d).
2. 2

   Reporting standards should be adopted.

   GUIDE-LLM covers roles, configurations, inputs, validation, reproducibility, and interests (Feuerriegel et al., 2026). Finance-specific work adds exact versions, query dates, prompts, generation settings, invalid response handling, and raw outputs (Coqueret et al., 2026; Wang et al., 2025c), along with temporal integrity, implementation costs, and internal agent disaggregation (Kong et al., 2026; Ye et al., 2026). Reporting enables assessment but does not validate simulations. Aggregate claims also require repeated markets and sessions.
3. 3

   For surrogate populations, individual histories offer a route to testable heterogeneity; reserve prices, survey answers, and public posts can anchor agents to identifiable human behavior (Bowles et al., 2026; Feng et al., 2025; Graham et al., 2026).

   Populating markets from matched human panels and comparing joint outcomes extends the same-design laboratory approach of Henning et al. (2025).
4. 4

   Dependence studies should measure covariance, marginal accuracy, and mean bias jointly, with explicit model and task populations and a justified independence baseline (Jo et al., 2026).

   Order-level herding statistics from empirical finance (Lakonishok et al., 1992; Sias, 2004) would make the action dependence of LLM agents comparable with that of institutional investors. Mixing families, post-training, or information can alter these aspects. Holding institutions fixed while measuring these changes would help separate dependence from individual errors and test algorithmic-monoculture predictions in financial settings (Kleinberg & Raghavan, 2021). Varying the share of each model within a market while recording every agent’s forecasts would also separate the initial disagreement from adaptation during the interaction (Saxena et al., 2026).
5. 5

   Contamination requires market outcome testing.

   Paired simulations using standard and time-restricted models (Benhenda, 2026; Yan et al., 2026) or real and counterfactual episodes (Bose, 2026) could test whether remembered outcomes affect simulated prices, provided capability, training, and information differences are held fixed or separately identified.
6. 6

   Institutions, assets, and the operating layer of deployed agents should be treated as experimental variables.

   Only four of the 30 price-setting designs vary the auction or trading mechanism (Chupilkin, 2026; Hansen & Lee, 2025; Shah et al., 2025b; Tsuchihashi, 2023). At least seven let agents trade more than one asset (Barton et al., 2026b; Gao et al., 2024; Li et al., 2026b; Lin, 2025; Ma et al., 2025; Yang et al., 2025; Zhang et al., 2026a), and several prohibit short selling (Lin, 2025; Lopez-Lira, 2025; Maresca, 2026). Real markets differ in price limits, auction formats, short-sale constraints, and market-making obligations, and these rules may interact with correlated LLM order flows in ways that single-asset laboratory designs cannot. The configuration surfaces of deployed agents are a further treatment. In one of the two deployed fleets, a risk slider sets leverage, and the list of assets displayed to the agent routes its selection (Barton et al., 2026b); thus, interface and configuration design are policy variables that experiments can manipulate.
7. 7

   Mixed human–LLM markets matter because they test substitution and participation.

   The corpus includes small mixed auctions: in one auction arena in the system audit, a human player bids against two LLM agents, and the highest bid sets each price (J. Chen et al., 2024). No study in the synthesis set runs a controlled asset market in which human and LLM traders jointly set prices, and Magar’s (2026) abstract describes a platform for such markets but reports no results from one. Human–machine games show that framing changes when humans join (Engel et al., 2025), whereas production venues contain other participants who are not characterized (Barton et al., 2026b). Varying the LLM share among human traders and extending the noise-trader design of Ross et al. (2026) would test both agent behavior and human responses.
8. 8

   Follow the advice through trades and prices.

   Model choice strongly shapes advice for a fixed household (Bäckman, 2026), and firms receiving buy recommendations have higher ESG scores after AI use is detected in the recommending analysts’ reports (Li & Yu, 2025). Linked platform data or field experiments could establish whether shared advice generates common exposures and has subsequent market effects.

From Section 9 of the review (read it in context). The numbers replace the paper’s “First, … Finally,”.