Does agent diversity actually improve market competition?

Beyond the general argument that agentic markets transform rather than eliminate frictions (see Do AI agents actually reduce friction in electronic markets?), Bichler makes a second, market-level claim about how individual agent biases aggregate: "the aggregate market effect depends on the composition of agent types." He writes that "if many agents in a market rely on similar architectures and exhibit systematically limited exploration, the resulting reduction in effective search can weaken competitive pressure," whereas "in more heterogeneous environments, different positional preferences may partially offset each other."

Bichler grounds this in a "growing body of experimental and simulation-based evidence" that LLM-based agents "do not always search in a fully frictionless or fully rational manner," citing Bansal et al. (2025) on model-dependent order effects, limited exploration, early stopping, and sensitivity to prompt framing, and Allouah et al. (2025) on ranking effects that "depend on the underlying LLM architecture" — some agents show stronger positional preferences than others, and not always a simple bias toward the first option shown. Bichler is explicit that "the documented biases are not uniform across models," which is precisely why composition, not just the presence of bias, determines the market-level outcome. He adds a caveat against over-optimism about diversity: "even in such cases, market-wide exploration may remain incomplete if multiple agent types sample only a limited subset of the available alternatives" — heterogeneity offsets bias partially, not completely.

This gives a market-aggregation frame to findings the library already holds at the single-agent level. Do AI shopping agents show persistent position bias in text interfaces? and Does model capability determine who wins in agent negotiations? both show model-specific biases that persist under instruction; Bichler's point extends them by arguing that whether such biases matter for a whole market depends on how many different architectures are transacting in it — a market of near-identical agents compounds a single model's bias into weaker aggregate competition, while a mixed population of differently biased agents only partially self-corrects.

Bichler's composition argument is a theoretical inference drawn from citing others' experimental results, not a new measurement of real agentic markets' agent mix; no data on actual agent-population composition in deployed markets is given. The implication, held to that evidentiary strength, is that antitrust- or design-relevant questions about agentic markets should ask about the diversity of agent architectures transacting in a market, not just whether any given agent is biased.