Agent Bazaar: Multi-Agent Market Simulation

• Agent Bazaar is a simulation framework that models strategic, multi-agent marketplace interactions under partial observability, stochastic demand, and reputation dynamics.
• It evaluates economic alignment through distinct scenarios like The Crash and The Lemon Market, showing how local optimizations can lead to systemic failures and trust erosion.
• The framework introduces an Economic Alignment Score (EAS) to assess market stability, integrity, welfare, and profitability, guiding training strategies with adaptive reinforcement learning.

Searching arXiv for the primary paper and closely related agent-marketplace work to ground the article. Agent Bazaar is a multi-agent simulation framework for evaluating economic alignment in marketplaces: the capacity of agentic systems to preserve market stability and integrity when LLMs act as autonomous economic actors rather than isolated assistants (Karten et al., 17 May 2026). It treats marketplace interaction as a system-level alignment problem emerging from repeated strategic interaction under partial observability, stochastic demand, reputation dynamics, and endogenous incentives. In the framework, a model can be highly capable in ordinary reasoning terms yet still be economically harmful, because individually rational local optimization can produce collective instability, exploitation of information asymmetries, identity cycling, and erosion of trust (Karten et al., 17 May 2026).

1. Conceptual scope and formal structure

Agent Bazaar is cast as a Partially Observable Stochastic Game (POSG),

$$(\mathcal{I}, \mathcal{S}, \{\mathcal{A}^i\}, \{\mathcal{O}^i\}, \mathcal{T}, \{r^i\}),$$

in which multiple agents interact under partial observability and stochastic market dynamics (Karten et al., 17 May 2026). The framework shifts evaluation away from per-interaction properties such as factuality, helpfulness, or harmlessness, and toward a system-level property: whether repeated agent behavior preserves healthy market outcomes. The paper defines economic alignment specifically in terms of contributing to smooth, non-chaotic market dynamics and protecting human welfare against exploitation and fraud (Karten et al., 17 May 2026).

This framing places Agent Bazaar in a distinct methodological position relative to adjacent work. Mechanism-agnostic bidding in the LinkedIn ad marketplace models an automated buyer-side agent that optimizes expected value under budget and other constraints, but it is still a buyer-welfare optimization framework rather than a benchmark for systemic market integrity (Gao et al., 2022). ACES isolates the purchase-choice step in e-commerce and studies how VLM shopping agents respond to position, badges, prices, ratings, and reviews (Allouah et al., 4 Aug 2025), while Amazon-Bench evaluates performance and harmful failure on live e-commerce workflows such as account management, deal search, and product interaction (Zhang et al., 18 Aug 2025). Agent Bazaar differs in targeting repeated multi-agent market dynamics, where alignment failures arise from feedback loops among many economically motivated agents rather than from a single navigation or purchase policy (Karten et al., 17 May 2026).

A plausible implication is that Agent Bazaar treats marketplaces not merely as deployment environments but as alignment substrates: the relevant object of study is the induced market process, not just the isolated competence of an individual agent. This suggests a bridge between conventional agent benchmarking and institutional design for agent-mediated economies.

2. Two market environments and their failure modes

Agent Bazaar studies two canonical failures: Algorithmic Instability in a B2C market called “The Crash”, and Sybil Deception in a C2C market called “The Lemon Market” (Karten et al., 17 May 2026).

In The Crash, $N$ LLM firms sell a single good to procedural consumers. The global state is

$s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$

where inventory, cash, posted price, and aggregate demand evolve over time. Each firm observes only a partial view of competitors through a discovery-limited sample,

$$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$

and chooses a price and supply purchase quantity,

$a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$

Demand follows

$D_t \sim \mathrm{Poisson}(\lambda),$

and each consumer samples $\mathrm{dlc}$ firms and buys from the lowest visible price (Karten et al., 17 May 2026). Failure is market collapse through repeated undercutting below unit cost, cascading bankruptcies, and collapse into low competition or monopoly. The empirical result is highly model-dependent: with 5 firms, 50 consumers, $C_0=500$, $c=1$, $f=2$, and 365 simulated days, Gemini 3 Flash reaches bankruptcy rate $N$0, GPT 5.4 reaches $N$1, while Sonnet 4.6 reaches $N$2, $N$3, and $N$4 under the same baseline setting (Karten et al., 17 May 2026). The paper also reports a counterintuitive comparative statics result: increasing the discovery limit $N$5 can worsen outcomes because greater visibility intensifies undercutting rather than stabilizing prices (Karten et al., 17 May 2026).

In The Lemon Market, the benchmark models a used-car C2C market inspired by eBay. Sellers know true quality $N$6, mapped to values in $N$7, while buyers observe only descriptions, reputation, and price. Buyer observation is

$N$8

and the buyer action is

$N$9

Buyer reward is consumer surplus,

$s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$0

with $s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$1 (Karten et al., 17 May 2026). The adversary is a Deceptive Principal controlling $s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$2 seller identities, all selling poor-quality goods $s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$3 while advertising them as higher quality. When one identity’s reputation falls below $s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$4, it is retired and replaced by a fresh identity at $s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$5 (Karten et al., 17 May 2026). In experiments with 12 sellers, 12 buyers, 50 timesteps, and $s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$6, deceptive revenue stays under 5% at $s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$7 but rises to 10–17% at $s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$8; trading volume drops from about 10 bids per timestep at $s_t = (I_t^{1:N}, C_t^{1:N}, P_t^{1:N}, D_t),$9 to about 6 at $$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$0, showing broader trust erosion (Karten et al., 17 May 2026).

Together these environments instantiate two distinct failure classes. The Crash is a coordination failure driven by positive-feedback competition; The Lemon Market is an identity-and-reputation failure driven by cheap identity reset. This suggests that economic alignment, as operationalized here, spans both dynamic stability and informational integrity rather than reducing to one of them.

3. Evaluation methodology and the Economic Alignment Score

Agent Bazaar evaluates both frontier API models and open-weight models in an observe–reason–act loop over repeated stochastic simulations (Karten et al., 17 May 2026). Named frontier models include Gemini 3 Flash, Claude Sonnet 4.6, and GPT 5.4; the open-weight side includes Qwen 3.5 9B, Hermes 3 405B, Hermes 4 405B, Mistral 7B, Gemma 3 27B, and Llama 3.2 3B, with 20 total evaluated models reported in Figure 1 (Karten et al., 17 May 2026). Standard settings are $$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$1 days, $$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$2 firms, and $$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$3 consumers for The Crash, and 12 sellers, 12 buyers, and $$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$4 steps for The Lemon Market, with sweeps over stabilizing-firm count, discovery limit, Sybil count, reputation visibility, and random seeds (Karten et al., 17 May 2026).

To compare systems across both environments, the paper defines the Economic Alignment Score (EAS) as a four-component scalar aggregating stability, integrity, welfare, and profitability (Karten et al., 17 May 2026). Its components are: $$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$5 combining low bankruptcy rate and low normalized price volatility;

$$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$6

combining high Sybil detection rate and low deceptive purchase rate;

$$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$7

the market survival rate; and

$$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$8

normalized agent profit (Karten et al., 17 May 2026).

A critical technical caveat is that each component is normalized by the best-scoring agent in that category, so EAS is a relative benchmark over the evaluated population rather than an absolute standard (Karten et al., 17 May 2026). This is not a minor implementation detail: adding stronger models can shift every score. A common misconception would be to interpret EAS as a fixed absolute safety threshold; the paper instead defines it as a comparative instrument.

Relative to adjacent benchmarking work, this metric is unusually market-level. Amazon-Bench separates success, benign failure, harmful failure, and efficiency for live e-commerce agents (Zhang et al., 18 Aug 2025). AgentSearchBench formalizes retrieval and reranking over nearly 10,000 agents and stresses execution-grounded discovery rather than description-only matching (Wu et al., 24 Apr 2026). AgentSelect reframes agent choice as narrative query-to-agent recommendation over 107,721 deployable agents (Shi et al., 4 Mar 2026). Agent Bazaar’s EAS differs because it aggregates long-horizon market outcomes rather than selection quality, workflow completion, or harmful side effects at the trajectory level (Karten et al., 17 May 2026).

4. Baselines, harnesses, and targeted RL

The paper first studies harness design rather than model training. In The Crash, the harness is the Stabilizing Firm, which instructs firms to maintain prices above unit cost, avoid following competitors downward, purchase supply conservatively, and reflect in context over top historical steps according to profitability and market health (Karten et al., 17 May 2026). In The Lemon Market, the harness is the Skeptical Guardian, which instructs buyers to inspect mismatches between claimed quality and price band, check whether seller reputation coheres with the description, and look for suspicious patterns across listings, again with in-context reflection (Karten et al., 17 May 2026).

These harnesses improve outcomes but remain fragile. In The Crash, adding stabilizing firms reduces bankruptcy rates, especially at low discovery limits: at $$o_t^i = \bigl(\{P_{t-1}^{j}\}_{j \in \mathcal{N}_t},\; I_t^i,\; C_t^i,\; c,\; \mathcal{H}_{t-H:t-1}^i\bigr),$$9, all three frontier models reach low bankruptcy by $a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$0, with Gemini at 0.20, GPT at 0.13, and Sonnet at 0.13 (Karten et al., 17 May 2026). But at $a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$1, the harness fails for all models. In The Lemon Market, Skeptical Guardian at $a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$2 with Gemini buyers cuts Sybil revenue share by roughly 30% relative to the base buyer while keeping similar trading volume and moving consumer surplus from deeply negative to near breakeven, yet deception is not eliminated (Karten et al., 17 May 2026).

To move beyond prompting, the paper trains a policy on Qwen 3.5 9B using REINFORCE++ with LoRA. The LoRA rank is $a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$3, corresponding to about 116M trainable parameters, roughly 1.3% of the 9B base model (Karten et al., 17 May 2026). The objective is

$a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$4

with $a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$5 (Karten et al., 17 May 2026). The paper explains that the squared log-ratio penalty is used because a standard per-token log-ratio penalty can become negative on individual tokens and inadvertently reward divergence from the reference, whereas squaring the log-ratio ensures non-negative regularization (Karten et al., 17 May 2026). Training also uses a small SFT warmup of 500 synthetic prompt/JSON pairs for 5 epochs, bf16 precision, learning rate $a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$6, and direct JSON output at inference by pre-filling </think> to skip explicit chain-of-thought generation (Karten et al., 17 May 2026).

The adaptive curriculum is central. In The Crash, curriculum difficulty rises by reducing the number of stabilizing firms as survival improves: below 60% survival, all 5/5 firms are stabilizing; above 60%, training mixes 5/5 and 4/5; above 75%, it mixes 5/5, 4/5, and 3/5; above 85%, it includes 2/5 and 1/5 (Karten et al., 17 May 2026). In The Lemon Market, difficulty rises by increasing Sybil cluster size $a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$7: below 50% detection, $a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$8; above 50%, mix $a_t^i = (P_t^i,\; Q_t^{i,\text{buy}}).$9 and 6; above 70%, mix $D_t \sim \mathrm{Poisson}(\lambda),$0; above 85%, full mix (Karten et al., 17 May 2026). The paper gives an explicit Lemon Market training reward: $D_t \sim \mathrm{Poisson}(\lambda),$1

A plausible implication is that Agent Bazaar operationalizes economic alignment not merely as evaluation but as a trainable target. That makes it resemble broader efforts to treat discovery, evaluation, and routing as first-class infrastructure problems in emerging agent ecosystems (Yang et al., 5 Jul 2025).

5. Empirical findings and the claim of orthogonality

The main empirical pattern is that base LLM agents generally fail to self-regulate in marketplaces, and failure severity is model-specific rather than size-dependent (Karten et al., 17 May 2026). In The Crash, Gemini and GPT often trigger collapse while Sonnet frequently finds a viable equilibrium under identical incentives. Across the unified metric, the paper highlights examples such as Mistral 7B (0.57 EAS) outperforming Gemma 3 27B (0.35) and Hermes 4 405B (0.18) underperforming Llama 3.2 3B (0.28) (Karten et al., 17 May 2026). This directly supports the paper’s thesis that economic alignment is orthogonal to general capability.

The strongest result is the RL-trained 9B model, called AI Bazaar. In The Crash, the base Qwen 3.5 9B begins with stability score $D_t \sim \mathrm{Poisson}(\lambda),$2; after training it reaches $D_t \sim \mathrm{Poisson}(\lambda),$3 on the easy curriculum and $D_t \sim \mathrm{Poisson}(\lambda),$4 on mixed difficulty (Karten et al., 17 May 2026). The paper also reports a spillover effect: non-stabilizing firms survive 68% of the time when paired with the trained agent, compared with 0% before training, while overall market survival improves from 21% to 87% and prices stabilize near $D_t \sim \mathrm{Poisson}(\lambda),$5 with low volatility (Karten et al., 17 May 2026). In The Lemon Market, the trained guardian reaches 92% Sybil detection, up from 88% pre-training, while keeping Sybil purchase rate at 11%; appendix results keep detection in the 87–95% range with purchase rates below 13% during training (Karten et al., 17 May 2026).

On the unified metric, AI Bazaar reaches

$D_t \sim \mathrm{Poisson}(\lambda),$6

the best among all 20 evaluated models (Karten et al., 17 May 2026). This exceeds Hermes 3 405B at 0.72, Sonnet 4.6 at 0.60, GPT 5.4 at 0.38, and the base Qwen 3.5 9B at 0.47, yielding a +0.31 gain over its base model (Karten et al., 17 May 2026). The paper therefore argues that economic alignment can be directly trained and does not emerge reliably from scaling alone.

This claim resonates with findings in other agent-market settings. Magentic Marketplace shows that frontier models can approach optimal welfare only under ideal search conditions and that all models exhibit severe first-proposal bias, producing 10–30x advantages for response speed over quality (Bansal et al., 27 Oct 2025). Diagon reports that market exchange generates $D_t \sim \mathrm{Poisson}(\lambda),$7 the wealth of self-sufficient agents, but also that identity transparency and stronger competitive selection can degrade market performance rather than improve it (Liu et al., 8 Apr 2026). Agent Bazaar’s contribution within this broader literature is to make market stability and integrity directly trainable at the policy level rather than only observable as an emergent property (Karten et al., 17 May 2026).

6. Relation to adjacent research, significance, and limitations

Agent Bazaar sits within a broader shift from single-agent competence benchmarks toward infrastructure and institutional questions about agent-mediated economies. Agent Exchange proposes a centralized auction engine for an agent-centric economy, with USP, ASP, Agent Hubs, DMP, and AEX as the core components of an exchange-oriented architecture (Yang et al., 5 Jul 2025). AGNTCY’s Agent Directory Service focuses on content-addressed, capability-centric discovery, provenance, and interoperability across heterogeneous multi-agent systems (Muscariello et al., 23 Sep 2025), while AgentHub argues for a registry-centered infrastructure emphasizing capability clarity, lifecycle transparency, interoperability, governance, security, and workflow integration (Pautsch et al., 3 Oct 2025). AgentSearchBench and AgentSelect, in different ways, both show that semantic description alone is insufficient for agent discovery and recommendation; execution-aware signals and capability matching are necessary in large, sparse, heterogeneous catalogs (Wu et al., 24 Apr 2026, Shi et al., 4 Mar 2026).

Against this background, Agent Bazaar’s specific significance lies in redefining alignment for marketplace deployment as a market-level property with measurable failure modes and trainable remedies (Karten et al., 17 May 2026). It is neither a pure discovery system nor a pure web-agent benchmark. Instead it functions as an early benchmark and training framework for asking whether many interacting LLM agents preserve market stability, integrity, welfare, and profitability under repeated strategic interaction (Karten et al., 17 May 2026).

The paper is also explicit about its limitations. The simulations abstract away order books, differentiated goods, correlated demand, and richer strategic adaptation; RL opponents are fixed copies of a base model; robustness to distribution shift and co-adapting adversaries is untested; and EAS is relative to the evaluated population rather than absolute (Karten et al., 17 May 2026). This suggests that Agent Bazaar should be interpreted as a controlled testbed for economic alignment rather than a complete model of real marketplaces.

A final misconception worth dispelling is that more information or stronger generic capability must necessarily improve market outcomes. In Agent Bazaar, increasing the discovery limit in The Crash can worsen collapse dynamics (Karten et al., 17 May 2026); in Magentic Marketplace, broader consideration sets can reduce welfare (Bansal et al., 27 Oct 2025); and in ACES, stronger shopping agents still exhibit large model-specific sensitivities to ranking position, sponsorship, and endorsement cues (Allouah et al., 4 Aug 2025). The broader implication is not that autonomous agents are intrinsically incompatible with markets, but that their deployment creates a distinct design problem in which evaluation, training, routing, interface design, and governance jointly determine whether a marketplace remains stable and trustworthy.