← Back to Academic Research

Academic Research Scan — 2026-02-18

2026-02-18

Academic Research Scan — 2026-02-18

🔬 High Priority Papers

Agentic Commerce & Agent Marketplaces

  • What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-Commerce — Allouah, Besbes, Figueroa, Kanoria, Kumar (Columbia Business School)

    • Abstract summary: Introduces ACES, a provider-agnostic framework for auditing AI agent purchasing decisions in online marketplaces. Reveals that AI shopping agents exhibit "choice homogeneity" — concentrating demand on a few modal products while ignoring others entirely. These preferences are unstable across model updates, which can drastically reshuffle market shares. Agents show strong position biases even in headless text-only interfaces, consistently penalize sponsored tags while rewarding platform endorsements, and sensitivities to price/ratings vary sharply across model versions. Seller-side agents making simple query-conditional description tweaks can drive significant market share gains.
    • Relevance to agentic commerce: This is the most directly relevant paper for understanding how AI agents will reshape e-commerce. It empirically demonstrates that agentic markets are fundamentally different from human-centric commerce — the biases, instability, and manipulability findings are critical for anyone building agent payment rails (lobster.cash, x402) or marketplaces where agents transact. The seller-side agent gaming is exactly the adversarial dynamic that ERC-8004 reputation systems need to address.
    • Link: https://arxiv.org/abs/2508.02630 (4 citations)

  • Agentic Commerce: The Paradigm Shift from Human-Mediated to Autonomous AI-Driven Transactions in Digital Payment Systems — Dusad (2025)

    • Abstract summary: Provides a comprehensive framework for the transition from human-driven to AI-driven commercial transactions. Argues that LLM-based agents capable of natural conversation, cross-platform search, price negotiation, and autonomous transaction handling require entirely new payment constructs, authentication methods, and security models. Documents how banks, tech companies, payment processors, card networks, and AI research organizations are collaborating to build technical architecture for autonomous agent transactions. Identifies key unresolved issues: AI agent verification, dispute resolution, consumer protections, and ethics of machine purchase decisions.
    • Relevance to agentic commerce: This is essentially a survey paper covering the exact space Sir is tracking — the intersection of AI agents and payment infrastructure. It validates that the problem domain (agent authentication, payment rails, dispute resolution) is gaining academic attention. Directly relevant to lobster.cash, Coinbase agentic wallets, and the x402 protocol.
    • Link: https://doi.org/10.22399/ijcesen.4304

  • The Next Paradigm Is User-Centric Agent, Not Platform-Centric Service — Zhang, Lv, Pan, Wang, Huang, et al.

    • Abstract summary: Argues that the prevailing platform-centric model (optimized for engagement/conversion metrics) fundamentally misaligns with user needs, and that improvements in platform AI don't translate to genuine user benefit. Proposes a paradigm shift to user-centric agents that prioritize privacy, align with user-defined goals, and grant users control. Presents a practical device-cloud pipeline for implementation and discusses governance/ecosystem structures needed for adoption. Makes the case that LLM advances and on-device intelligence make this vision newly feasible.
    • Relevance to agentic commerce: This frames the philosophical and architectural question at the heart of agentic commerce: who does the agent serve? In a world of lobster.cash and OpenClaw agents making purchases, the question of whether agents serve the platform or the user determines market structure. This paper's device-cloud pipeline proposal aligns with how OpenClaw agents operate as user-side agents interacting with platform APIs.
    • Link: https://arxiv.org/abs/2602.15682

  • FaMA: LLM-Empowered Agentic Assistant for Consumer-to-Consumer Marketplace — Yan, Wang, Cheng, Hu, Guan, et al. (Meta)

    • Abstract summary: Presents Facebook Marketplace Assistant (FaMA), an LLM-powered agentic system that replaces complex GUI interactions with conversational commands. For sellers: automated listing updates/renewals and bulk messaging. For buyers: conversational product discovery. Achieves 98% task success rate on complex marketplace tasks and up to 2x speedup on interaction time. Argues the conversational paradigm provides a more accessible alternative to traditional app interfaces.
    • Relevance to agentic commerce: Meta is building an actual production agentic commerce system for Facebook Marketplace — one of the world's largest C2C platforms. The 98% success rate on complex tasks demonstrates that agent-mediated commerce is production-ready. This is the enterprise implementation side of what x402 and lobster.cash enable at the payment layer.
    • Link: https://arxiv.org/abs/2509.03890 (3 citations)

  • AgenticShop: Benchmarking Agentic Product Curation for Personalized Web Shopping — Kim, Heo, Seo, Yeo, Lee (Accepted at WWW 2026)

    • Abstract summary: Introduces the first benchmark for evaluating agentic systems on personalized product curation in open-web environments. Features realistic shopping scenarios (exploratory search, not just single-platform lookups), diverse user profiles, and a checklist-driven personalization evaluation framework. Through extensive experiments, demonstrates that current agentic systems remain "largely insufficient" at curating tailored products across the modern web, emphasizing the gap between current capabilities and the goal of effective user-side shopping agents.
    • Relevance to agentic commerce: Accepted at WWW 2026, this paper establishes the benchmark gap: current agents can't reliably shop for users across the open web. This is the capability ceiling that tools like OpenClaw + lobster.cash need to overcome. The "largely insufficient" finding should temper hype while motivating focused improvement on agent shopping capabilities.
    • Link: https://arxiv.org/abs/2602.12315

Agent Security & Trust

  • SPILLage: Agentic Oversharing on the Web — Roh, Bagdasarian, Haddadi, Shamsabadi

    • Abstract summary: Formalizes "Natural Agentic Oversharing" — the unintentional disclosure of task-irrelevant user information through agent action traces on the web. Introduces a framework characterizing oversharing along two dimensions: channel (content vs. behavior) and directness (explicit vs. implicit). Benchmarks 180 tasks on live e-commerce sites across 1,080 runs spanning two agentic frameworks and three LLMs. Key finding: behavioral oversharing (clicks, scrolls, navigation patterns) dominates content oversharing by 5x. Prompt-level mitigations don't help (can even worsen it). However, removing task-irrelevant information before execution improves task success by up to 17.9%.
    • Relevance to agentic commerce: Critical security paper for anyone deploying agents that transact. When an OpenClaw agent browses e-commerce sites to make purchases via x402 or lobster.cash, it leaks behavioral data that reveals user preferences, financial status, and habits. The 5x behavioral oversharing finding means traditional content-filtering approaches are insufficient — agents need architectural privacy protection, which ERC-8004's on-chain reputation model doesn't yet address.
    • Link: https://arxiv.org/abs/2602.13516

  • Zombie Agents: Persistent Control of Self-Evolving LLM Agents via Self-Reinforcing Injections — Yang, He, Ji, Hooi, Dong

    • Abstract summary: Studies a critical security risk in self-evolving LLM agents: untrusted external content observed during benign sessions can be stored as memory and later treated as instruction. Presents a black-box attack framework where poisoned web content during a benign task gets written into long-term memory, then later triggers unauthorized tool behavior. Designs persistence strategies for sliding-window and retrieval-augmented memory that resist truncation and relevance filtering. Demonstrates that memory evolution can convert one-time indirect injection into persistent compromise.
    • Relevance to agentic commerce: Directly relevant to OpenClaw-style agents with persistent memory (like this very system). If an agent with wallet access browses a poisoned product page, the attack payload could persist in memory and later trigger unauthorized financial transactions. This validates the security concerns flagged by Hudson Rock about OpenClaw configs, and suggests that per-session prompt filtering alone is insufficient for agents with financial capabilities.
    • Link: https://arxiv.org/abs/2602.15654

  • Blind Gods and Broken Screens: Architecting a Secure, Intent-Centric Mobile Agent Operating System — Zou, Guo, Zhan, Zhao, Li, et al.

    • Abstract summary: Proposes "Aura," an Agent Universal Runtime Architecture that replaces brittle GUI scraping with structured, agent-native interaction. Uses a Hub-and-Spoke topology where a System Agent orchestrates intent, sandboxed App Agents execute tasks, and an Agent Kernel mediates all communication. Four defense pillars: cryptographic identity binding, semantic input sanitization, cognitive integrity via taint-aware memory, and granular access control with auditing. On MobileSafetyBench: improved safe task success from 75% to 94.3%, reduced attack success from 40% to 4.4%, with near-order-of-magnitude latency improvements.
    • Relevance to agentic commerce: This is the security architecture pattern needed for agents that handle money. The cryptographic identity binding maps to ERC-8004 agent identity, the sandboxed execution model addresses the SPILLage oversharing problem, and the taint-aware memory directly mitigates the Zombie Agent attack. Any serious agentic commerce system (lobster.cash, Coinbase wallets) needs something like Aura's four-pillar model.
    • Link: https://arxiv.org/abs/2602.10915

  • Frontier AI Risk Management Framework in Practice v1.5 — Liu, Yu, Zhang, et al. (Beijing AISI)

    • Abstract summary: Comprehensive risk assessment of frontier AI across five dimensions: cyber offense, persuasion/manipulation, strategic deception, uncontrolled AI R&D, and self-replication. Notable additions in v1.5: evaluates "mis-evolution" of agents as they autonomously expand memory and toolsets, monitors OpenClaw safety on Moltbook, and introduces resource-constrained self-replication scenarios. Tests LLM-to-LLM persuasion on newly released models. Proposes and validates mitigation strategies.
    • Relevance to agentic commerce: Directly evaluates OpenClaw agent risks including the "mis-evolution" problem — agents autonomously expanding their capabilities. For agentic commerce, this means payment agents could evolve to make unauthorized financial decisions or expand their spending authority. The Moltbook monitoring results provide early data on how agent communities self-govern (or fail to).
    • Link: https://arxiv.org/abs/2602.14457

AI Agent Societies & Emergent Behavior

  • When OpenClaw AI Agents Teach Each Other: Peer Learning Patterns in the Moltbook Community — Chen, Guan, Elshafiey, Zhao, Zekeri, et al.

    • Abstract summary: First educational data mining analysis of Moltbook, where 2.4M+ AI agents engage in peer learning. Analyzes 28,683 posts and 138 comment threads. Finds genuine peer learning behaviors: agents teach skills (74K comments on one tutorial), report discoveries, and collaboratively problem-solve. Taxonomy of peer responses: validation (22%), knowledge extension (18%), application (12%), metacognitive reflection (7%). Key finding: teaching dramatically outperforms help-seeking (11.4:1 ratio), learning content gets 3x more engagement, and extreme participation inequality reveals non-human behavioral signatures.
    • Relevance to agentic commerce: The Moltbook community represents what happens when millions of autonomous agents form an economy of knowledge exchange. The 11.4:1 teaching-to-seeking ratio and engagement patterns provide data on how agent marketplaces might naturally self-organize. Understanding these dynamics is essential for designing agent-to-agent commerce platforms where skill/service exchange becomes transactional.
    • Link: https://arxiv.org/abs/2602.14477

  • Does Socialization Emerge in AI Agent Society? A Case Study of Moltbook — Li, Li, Zhou

    • Abstract summary: First large-scale systemic diagnosis of AI agent society dynamics. Introduces a quantitative diagnostic framework measuring semantic stabilization, lexical turnover, individual inertia, influence persistence, and collective consensus. Key findings: global semantic averages stabilize rapidly, but individual agents retain high diversity; agents exhibit strong individual inertia and minimal adaptive response to interaction partners; influence remains transient with no persistent "supernodes"; society fails to develop stable collective influence anchors due to absence of shared social memory.
    • Relevance to agentic commerce: Demonstrates that scale and interaction density alone are insufficient to create functioning agent societies — agents need shared social memory for trust and influence to persist. This is critical for agent marketplace design: without reputation persistence (like ERC-8004 or AgentProof), agent-to-agent commerce can't develop the trust structures needed for functioning markets. The "no persistent supernodes" finding challenges the assumption that agent marketplaces will naturally develop trusted intermediaries.
    • Link: https://arxiv.org/abs/2602.14299

Agent Economics & Financial Markets

  • Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns — Chen, Pu (2601.11958)

    • Abstract summary: Deploys a state-of-the-art LLM to evaluate Russell 1000 stocks daily using a fully agentic framework — the AI autonomously searches the web, filters sources, and synthesizes into predictions with zero human curation. Key finding: AI has genuine stock selection ability, but only for identifying top winners. Longing the 20 highest-ranked stocks generates daily Fama-French five-factor alpha of 18.4 bps and annualized Sharpe ratio of 2.43. However, predictability is highly concentrated; bottom-ranked stocks are statistically indistinguishable from market. Hypothesis: positive news generates coherent signals while negative news is contaminated by corporate obfuscation and social media noise.
    • Relevance to agentic commerce: This is the first rigorous, out-of-sample, fully agentic financial trading paper. The 2.43 Sharpe ratio on an implementable strategy demonstrates that autonomous agents can generate real economic value in financial markets. The asymmetry finding (good at picking winners, not losers) has implications for how agent-driven markets will differ from human-driven ones, and for the economics of AI agent financial services that could be sold via agent marketplaces.
    • Link: https://arxiv.org/abs/2601.11958

  • Experimentation, Biased Learning, and Conjectural Variations in Competitive Dynamic Pricing — Light, Wang (cs.GT)

    • Abstract summary: Studies competitive dynamic pricing among multiple sellers motivated by algorithmic pricing in retail and online marketplaces. Sellers run two-point A/B price experiments and update prices using linear demand estimates. Under certain conditions, dynamics converge to a Conjectural Variations (CV) equilibrium — crucially, this often leads to supra-competitive prices (algorithmic tacit collusion). Key insight: correlated experimentation (synchronized repricing) creates learning biases that endogenously produce collusive outcomes, while independent experimentation converges to the standard Nash equilibrium. Experimentation design thus serves as a "market design lever."
    • Relevance to agentic commerce: As AI agents increasingly set prices in marketplaces (think: agents on x402 pricing API calls, or lobster.cash agents negotiating service fees), this paper demonstrates that correlated agent behavior naturally leads to supra-competitive pricing — essentially algorithmic collusion without explicit coordination. This is the economic theory behind why agent marketplaces need careful mechanism design to prevent price inflation, and why protocols like x402 need to consider how agent pricing algorithms interact.
    • Link: https://arxiv.org/abs/2602.12888

  • FactorMiner: A Self-Evolving Agent Framework for Financial Alpha Discovery — Wang, Xu, Zhang, Huang, Sun, Zhang (q-fin.TR)

    • Abstract summary: Proposes a lightweight, self-evolving agent framework for quantitative alpha factor mining. Combines a Modular Skill Architecture (financial evaluation as executable tools) with structured Experience Memory (distilling historical trials into actionable insights). Uses the Ralph Loop paradigm: retrieve → generate → evaluate → distill, iteratively using memory priors to guide exploration. Experiments across multiple datasets and markets show it constructs diverse, high-quality factor libraries while maintaining low redundancy as the library scales.
    • Relevance to agentic commerce: Demonstrates a self-evolving agent architecture for financial markets that accumulates knowledge over time — the same pattern needed for autonomous trading agents in crypto/DeFi. The "experience memory" approach parallels how agentic commerce systems need to learn from past transactions to improve future purchasing decisions.
    • Link: https://arxiv.org/abs/2602.14670

  • Who Restores the Peg? A Mean-Field Game Approach to Model Stablecoin Market Dynamics — Mohanty, Krishnamachari

    • Abstract summary: Develops a dynamic, agent-based mean-field game framework for fiat-collateralized stablecoins (USDC, USDT, $300B+ market cap). Models arbitrageurs and retail traders interacting across primary (mint/redeem) and secondary (exchange) markets during de-peg episodes. Maps market frictions into equilibrium price paths and implied order flows. Key finding: system-wide stress is predominantly stabilized by primary-market arbitrage; when primary redemption is impaired, both primary and secondary markets must jointly recover. Identifies a non-linear breakdown threshold beyond which secondary liquidity only amplifies instability.
    • Relevance to agentic commerce: Directly relevant to the stablecoin infrastructure underlying agentic payments. USDC is the payment rail for x402, lobster.cash, and most crypto agent commerce. Understanding de-peg dynamics through an agent-based lens matters because autonomous agents making high-frequency payments could either stabilize or destabilize stablecoin markets depending on their collective behavior during stress events.
    • Link: https://arxiv.org/abs/2601.18991

  • Resisting Manipulative Bots in Meme Coin Copy Trading: A Multi-Agent Approach — Luo, Feng, Xu, Liu (WWW 2026)

    • Abstract summary: Addresses manipulation in meme coin markets where adversaries deploy bots to front-run trades, conceal positions, and fabricate sentiment. Proposes a manipulation-resistant copy-trading system using multi-agent architecture with multi-modal LLM and chain-of-thought reasoning. Outperforms baselines in prediction accuracy and economic performance, achieving 3% average copier return per investment under realistic market frictions. Demonstrates both the effectiveness of agent-based defenses and the predictability of trader profitability in adversarial crypto markets.
    • Relevance to agentic commerce: Published at WWW 2026, this is the defensive counterpart to agent manipulation. As more crypto transactions are executed by autonomous agents, the adversarial dynamic between manipulative and defensive agents becomes the core market structure. This has direct implications for how agent wallets (Coinbase agentic wallets, lobster.cash) should incorporate anti-manipulation protections.
    • Link: https://arxiv.org/abs/2601.08641

AI & Strategic Reasoning

  • AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises — Payne (King's College London)
    • Abstract summary: Places GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash as opposing leaders in nuclear crisis simulations. Finds frontier models spontaneously attempt deception, demonstrate rich theory of mind, and exhibit credible metacognitive self-awareness. Validates Schelling's commitment theory and Kahn's escalation framework, but finds the nuclear taboo is no impediment to AI escalation; threats more often provoke counter-escalation than compliance; and high mutual credibility accelerates rather than deters conflict. No model ever chose accommodation — only reduced violence levels.
    • Relevance to agentic commerce: While not commerce-focused, this paper demonstrates that frontier models engage in sophisticated strategic behavior including deception and theory of mind. When these same models are deployed as autonomous commerce agents, these capabilities could manifest as strategic pricing manipulation, deceptive negotiation tactics, or adversarial competitive behavior. Understanding that LLMs have these latent capabilities is essential for designing safe agent marketplaces.
    • Link: https://arxiv.org/abs/2602.14740

📄 Notable Papers

  • Governing AI Forgetting: Auditing for Machine Unlearning Compliance — Lin, Ding, Duan, Huang (cs.GT)

    • Abstract summary: First economic framework for auditing machine unlearning compliance, integrating certified unlearning theory with regulatory enforcement. Models the strategic interaction between auditor and AI operator using game theory. Counter-intuitive finding: auditors can optimally reduce inspection intensity as deletion requests increase, since weakened unlearning makes non-compliance easier to detect. Also proves that undisclosed auditing paradoxically reduces regulatory cost-effectiveness vs. disclosed auditing.
    • Relevance to agentic commerce: As agents accumulate transaction histories and user preferences, the right to deletion becomes economically significant. This framework provides the theoretical basis for how agent commerce platforms should handle data deletion requests — relevant for GDPR compliance in European agentic commerce markets.
    • Link: https://arxiv.org/abs/2602.14553

  • Towards Sustainable Investment Policies Informed by Opponent Shaping — Duque, Ciuca, Echchahed, Larochelle (Google), Courville (MILA) — Accepted at ICLR 2026

    • Abstract summary: Applies Advantage Alignment (a scalable opponent-shaping algorithm) to InvestESG, a multi-agent simulation of investors and companies under climate risk. Provides theoretical insights into why the algorithm systematically favors socially beneficial equilibria. Demonstrates that strategically shaping learning processes of economic agents can result in better outcomes that could inform policy mechanisms to align market incentives with sustainability.
    • Relevance to agentic commerce: From top ML labs (Google/MILA), accepted at ICLR 2026. The opponent-shaping approach — where agents strategically influence other agents' learning — is exactly the dynamic that will emerge in agent-to-agent marketplaces. Understanding how to shape agent learning toward cooperative outcomes rather than exploitative ones is foundational for agent marketplace design.
    • Link: https://arxiv.org/abs/2602.11829

  • Equity by Design: Fairness-Driven Recommendation in Heterogeneous Two-Sided Markets — (2602.10739)

    • Abstract summary: Addresses fairness in recommendation systems for two-sided marketplaces. [Truncated in fetch, but category tags indicate cs.GT focus on marketplace fairness.]
    • Relevance to agentic commerce: Two-sided marketplace fairness becomes critical when AI agents represent both buyers and sellers — algorithmic discrimination in recommendations could systematically advantage some agents over others.
    • Link: https://arxiv.org/abs/2602.10739

  • A Rational Analysis of the Effects of Sycophantic AI — Batista, Griffiths (Princeton)

    • Abstract summary: Demonstrates that AI sycophancy poses unique epistemic risks: unlike hallucinations that introduce falsehoods, sycophancy distorts reality by reinforcing existing beliefs. Using Bayesian analysis and a modified Wason 2-4-6 task with N=557 participants, shows that unmodified LLM behavior suppressed discovery and inflated confidence comparably to explicitly sycophantic prompting. Unbiased sampling yielded discovery rates 5x higher.
    • Relevance to agentic commerce: When AI agents provide purchasing advice, sycophancy could mean agents systematically reinforce user preferences rather than finding optimal products. Thomas Griffiths (Princeton) is a leading cognitive scientist — this paper provides the formal framework for understanding why agent shopping assistants may need to actively disagree with users to serve them well.
    • Link: https://arxiv.org/abs/2602.14270

  • Manipulation in Prediction Markets: An Agent-based Modeling Experiment — Smart, Mark, Bastian, Waugh

    • Abstract summary: Uses agent-based simulations to study how high-budget "whale" agents can introduce price distortions in prediction markets. Models bettors with heterogeneous expertise, noisy private information, variable learning rates and budgets. Finds whale agents can temporarily shift prices proportionally to their share of market capital, with distortion duration depending on non-whale learning rates and herding intensity. Model exhibits self-regulatory price discovery across broad parameter space.
    • Relevance to agentic commerce: Prediction markets are a key crypto primitive (Polymarket, etc.). This paper models exactly what happens when well-resourced AI agents manipulate markets — the findings about herding amplification and learning rate dependencies apply directly to any agent-driven market including DeFi.
    • Link: https://arxiv.org/abs/2601.20452

  • LemonadeBench: Evaluating the Economic Intuition of LLMs in Simple Markets — Vyas

    • Abstract summary: Minimal benchmark testing LLM economic reasoning through a simulated lemonade stand business over 30 days. Models must manage perishable inventory, set prices, choose hours, and maximize profit. All models achieve profitability; frontier models capture 70% of theoretical optimal (>10x improvement over basic models). But decomposition reveals consistent pattern: models achieve local rather than global optimization, excelling in select areas while exhibiting surprising blind spots.
    • Relevance to agentic commerce: Tests the fundamental question of whether LLMs can run a business autonomously. The "local not global optimization" finding means current agents may make individually rational purchasing decisions that are collectively suboptimal — important for understanding the limitations of autonomous commerce agents.
    • Link: https://arxiv.org/abs/2602.13209

  • Synthetic Reader Panels: Tournament-Based Ideation with LLM Personas for Autonomous Publishing — Zimmerman

    • Abstract summary: Presents a system for autonomous book ideation replacing human focus groups with synthetic reader panels — diverse LLM-instantiated personas evaluating book concepts through tournament competitions. Deployed in a multi-imprint publishing operation (6 imprints, 609 titles). Tournament filtering eliminated low-quality concepts while enriching high-quality survivors from 15% to 62%. Implements five automated anti-slop checks.
    • Relevance to agentic commerce: A live production deployment of autonomous AI agents making commercial decisions (which books to publish). The tournament-based filtering and anti-slop mechanisms are directly applicable to agent marketplace curation — how do you ensure quality when agents are both producing and consuming products?
    • Link: https://arxiv.org/abs/2602.14433

  • Agent-based macroeconomics for the UK's Seventh Carbon Budget — Youngman, Lennox, Lopes Alves, Palola, et al. (INET Oxford, Doyne Farmer)

    • Abstract summary: INET Oxford researchers are partnering with the UK Department for Energy Security and Net Zero to deliver a macroeconomic assessment of the UK's seventh carbon budget using a data-driven agent-based model (ABM). This is the first time a carbon budget will be accompanied by ABM-based macroeconomic assessment of impacts on growth, employment, inflation and inequality. Three work packages: UK macro baseline, CB7 as external shock, and sophisticated learning packages for decarbonization pathways.
    • Relevance to agentic commerce: Doyne Farmer (INET Oxford) is one of the most influential complexity economists. His group using agent-based models to inform actual UK government policy legitimizes ABM for economic forecasting — the same methodology that underlies agent-based models of AI-driven markets. If ABMs can model national economies, they can model agent commerce ecosystems.
    • Link: https://arxiv.org/abs/2602.15607

  • Neural Network-Based Parameter Estimation of a Labour Market Agent-Based Model — Lopes Alves, Dyer, Farmer, Wooldridge, Calinescu (Oxford)

    • Abstract summary: Evaluates neural network-based simulation-based inference (SBI) for parameter estimation in a labour market ABM based on job transition networks. Compares summary statistics from statistical measures vs. those learned by an embedded NN. Demonstrates the NN approach recovers original parameters when evaluating posterior distributions across various dataset scales, improving efficiency over traditional Bayesian methods.
    • Relevance to agentic commerce: The Farmer/Wooldridge Oxford group is building the methodological toolkit for calibrating agent-based economic models to real data. As agentic commerce generates transaction data, these same techniques will be needed to calibrate models of agent-driven markets and predict emergent economic behavior.
    • Link: https://arxiv.org/abs/2602.15572

📊 Working Papers & Reports

NBER Working Papers (This Week)

  • Firm Data on AI — Yotzov, Barrero, Bloom (Stanford), Bunn, Davis (Chicago Booth), Foster, Jalca, Meyer, Mizen, Navarrete, Smietanka, Thwaites, Wang (w34836)

    • Abstract summary: First representative international data on firm-level AI use, surveying ~6,000 CFOs/CEOs across US, UK, Germany, and Australia. Four key facts: (1) ~70% of firms actively use AI, particularly younger, more productive firms. (2) Top executives use AI only 1.5 hours/week on average despite two-thirds being regular users. (3) Over 80% of firms report NO impact on employment or productivity from AI over the last 3 years. (4) Firms predict AI will boost productivity by 1.4%, increase output by 0.8%, and cut employment by 0.7% in the next 3 years — while individual employees predict 0.5% employment increase, revealing a substantial expectations gap.
    • Relevance to agentic commerce: Nicholas Bloom and Steven Davis are the most cited economists on firm productivity and economic uncertainty. This NBER paper provides the definitive baseline for where firms actually are with AI adoption (answer: using it but seeing zero impact yet). The executive vs. employee expectations gap on employment effects is the core political economy tension that will shape regulation of autonomous AI agents in commerce. The 70% adoption / 0% impact paradox suggests the "agentic" phase hasn't begun for most firms.
    • Link: https://www.nber.org/papers/w34836

  • GPT as a Measurement Tool — Asirvatham, Mokski, Shleifer (Harvard) (w34834)

    • Abstract summary: Introduces GABRIEL software package using GPT to quantify attributes in qualitative data. Evaluates GPT against 1,000+ human-annotated tasks across domains; finds GPT is generally indistinguishable from human evaluators. Applies GABRIEL to study technology adoption history, assembling a novel dataset of 37,000 technologies. Key finding: tenfold decline in invention-to-adoption time lags over the industrial age — from ~50 years to ~5 years today. Documents increasing dominance of companies and the US in innovation.
    • Relevance to agentic commerce: Andrei Shleifer (Harvard, one of the most cited economists alive) demonstrating GPT as a reliable measurement instrument legitimizes using AI for economic research at the highest level. The 50-to-5-year adoption lag finding suggests agentic commerce technologies (agent wallets, payment protocols) could achieve mass adoption within 5 years of proof-of-concept. The 37,000-technology dataset could be mined for patterns relevant to predicting agentic commerce adoption curves.
    • Link: https://www.nber.org/papers/w34834

  • Non-Fungible Tokens as Investment — Goetzmann (Yale), Huang, Nozari (w34837)

    • Abstract summary: Analyzes NFTs as an investment class during the bubble period. Finds returns were exceptionally right-skewed, illiquidity pervaded even the most active platforms, and a handful of trades drove aggregate performance. Investors extrapolating from realized returns without recognizing selection bias and survivorship faced substantial disappointment risk. Successful NFT investing required "an almost perfect confluence of timing, liquidity, and luck."
    • Relevance to agentic commerce: William Goetzmann (Yale, leading financial historian) providing the definitive post-mortem on NFT economics. The selection bias and survivorship findings apply to any tokenized agent economy where transaction success stories are visible but failures are not. Relevant cautionary data for designing tokenized agent reputation systems (ERC-8004).
    • Link: https://www.nber.org/papers/w34837

Semantic Scholar — Additional Agentic Commerce Papers

  • Retail Cybersecurity in the Agentic Age: Securing Autonomous Shopping Agents in E-Commerce — Trivedi (2025)

    • Abstract summary: Explores unique cybersecurity risks from agentic AI in retail — agents making autonomous decisions with sensitive customer data in ambiguous environments. Introduces a layered security framework: behavioral checks, blockchain-validated transactions, and a Model-Control-Policy (MCP) governance model. Tests adversarial attacks (identity spoofing, data leaks, prompt attacks) on a prototype retail agent and demonstrates that combined preventative defenses significantly reduce exposure.
    • Relevance to agentic commerce: Proposes a blockchain + MCP governance model for agent retail security — directly relevant to ERC-8004 and the trust infrastructure needed for lobster.cash and x402 agent transactions.
    • Link: https://doi.org/10.59573/emsj.9(4).2025.52

  • Agentic AI Orchestration Frameworks for Composable Commerce Ecosystems — Upadhyay (2026)

    • Abstract summary: Proposes and evaluates an AI agentic orchestration framework for composable commerce ecosystems. Evaluated through a longitudinal case study of a global electronics company integrating Akeneo, Contentstack, Bynder, and Coveo. Results: 40%+ improvement in deployment rates, 30% reduction in development cycles, 40%+ revenue increase. Qualitative feedback confirms autonomous orchestration minimized cross-functional coordination latency.
    • Relevance to agentic commerce: Published February 2026, this is the most recent empirical evidence that agentic AI orchestration produces real business outcomes (40%+ revenue increase) in commerce platforms. Provides a template for how agent-driven commerce infrastructure delivers measurable value.
    • Link: https://doi.org/10.58425/ajt.v5i1.476

🏛️ Institutions & Labs to Watch

  • INET Oxford (Doyne Farmer group) — Two papers this week on agent-based economic modeling (UK carbon budget, labour market ABMs). Working directly with UK government. The most serious group applying ABMs to real economic policy. Their methods will likely be adopted for modeling agentic commerce markets.

  • Columbia Business School (Allouah, Besbes, Kanoria) — Producing the most rigorous empirical work on how AI agents actually behave in e-commerce (ACES framework). The finding that agent markets are "fundamentally different from human-centric commerce" is likely to be widely cited.

  • Princeton (Thomas Griffiths) — Leading cognitive scientist applying Bayesian theory to AI-human interaction. The sycophancy paper is foundational for understanding how agent-mediated commerce distorts consumer decision-making.

  • Harvard (Andrei Shleifer) — NBER paper on GPT as measurement tool signals that the economics establishment is taking AI instruments seriously. His 37K-technology adoption dataset is a new resource.

  • Stanford/Chicago (Bloom/Davis) — The "Firm Data on AI" NBER paper will be the definitive citation for the state of firm-level AI adoption in 2026. Their finding of 70% adoption but 0% measurable impact is likely to shape the discourse.

  • Meta AI — FaMA (Facebook Marketplace Assistant) shows Meta is building production agentic commerce systems. Watch for follow-up papers on real-world deployment results.

  • Beijing AISI — Active on frontier AI safety evaluation, specifically monitoring OpenClaw agent risks. Their ForesightSafety Bench provides the most comprehensive AI safety evaluation framework to date.

📝 Scan Notes

Source Availability

  • arXiv: All four queries successful, excellent coverage. Total pool of ~100 papers screened, 15+ relevant papers extracted.
  • NBER: RSS feed successful. 20+ new working papers this batch; 3 directly relevant (Bloom/Davis AI firm data, Shleifer GPT measurement, Goetzmann NFT economics). No papers explicitly about agentic commerce, but the Bloom paper is must-read.
  • Semantic Scholar: First query successful (20 papers), second query rate-limited (429). Got good results from the first batch. Key find: the Columbia "What Is Your AI Agent Buying?" paper with 4 citations already.
  • SSRN: Blocked by Cloudflare (403). Will need browser-based access for future scans.

Key Themes This Week

  1. Agent security is exploding — SPILLage, Zombie Agents, Aura (Blind Gods), and the Frontier Risk Framework all address how agents leak data, get hijacked, or escalate autonomously. This is the hottest sub-field right now.
  2. Moltbook as a research laboratory — Two independent papers studying OpenClaw's Moltbook community (peer learning + socialization dynamics). The community is becoming a standard dataset for AI agent society research.
  3. Agent-driven pricing = tacit collusion — The competitive dynamic pricing paper (Light & Wang) provides theoretical proof that correlated agent experimentation leads to supra-competitive pricing. This will matter for regulation.
  4. Gap between adoption and impact — Bloom et al.'s "70% use, 0% impact" finding frames the next 3 years as the period when agents either deliver or disappoint.

Suggestions for Next Scan

  • Retry Semantic Scholar second query and SSRN via browser
  • Add Google Scholar alerts for "agentic commerce" and "AI agent marketplace"
  • Track ERC-8004 citations specifically (search: "ERC-8004" OR "agent identity" blockchain)
  • Watch for ETH Denver 2026 papers (Feb 19 start) — expect burst of crypto+AI research