← Back to Academic Research

Academic Research Scan — 2026-02-17

2026-02-17

Academic Research Scan — 2026-02-17

🔬 High Priority Papers

Agentic Commerce & Payments

  • Zero-Trust Runtime Verification for Agentic Payment Protocols: Mitigating Replay and Context-Binding Failures in AP2 — Qianlong Lan, Anuj Kaul, Shaun Jones, Stephanie Westrum

    • Abstract summary: Presents a security analysis of the AP2 (Agent Payments Protocol) mandate lifecycle, identifying enforcement gaps when AI agents execute payments in real-world conditions with retries, concurrency, and orchestration. The authors propose a zero-trust runtime verification framework using dynamically generated time-bound nonces to enforce consume-once mandate semantics. Simulation-based evaluation shows the framework mitigates all evaluated replay and context-redirect attacks while maintaining ~3.8ms verification latency at up to 10,000 TPS. Key insight: runtime state is bounded by peak concurrency, not cumulative history.
    • Relevance to agentic commerce: Directly addresses the security layer that protocols like AP2, x402, and ERC-8004 need for production deployment. The concurrency/replay attack vectors identified are exactly the kind of vulnerabilities that emerge when agents like OpenClaw autonomously execute purchases. Shows that robust agentic payments are achievable with minimal overhead.
    • Link: https://arxiv.org/abs/2602.06345

  • Whispers of Wealth: Red-Teaming Google's Agent Payments Protocol via Prompt Injection — Tanusree Debi, Wentian Zhu

    • Abstract summary: Performs AI red-teaming of Google's AP2 protocol and discovers critical vulnerabilities through indirect and direct prompt injection. Introduces two novel attacks: the "Branded Whisper Attack" (manipulates product ranking in agent shopping) and the "Vault Whisper Attack" (extracts sensitive user data). Using a functional AP2 shopping agent built with Gemini-2.5-Flash and Google ADK, they experimentally validate that simple adversarial prompts can reliably subvert agent purchasing behavior. Demonstrates that cryptographic verification alone is insufficient—LLM reasoning must also be secured.
    • Relevance to agentic commerce: Critical finding for anyone building agentic shopping (lobster.cash, OpenClaw commerce). Shows that even well-designed payment protocols like AP2 can be subverted through the LLM layer. Implies that x402's simpler model (no LLM reasoning in payment loop) may actually be more secure for certain use cases.
    • Link: https://arxiv.org/abs/2601.22569

  • Know Your Contract: Extending eIDAS Trust into Public Blockchains — Awid Vaziry, Christoph Wronka, Sandro Rodriguez Garzon, Axel Küpper

    • Abstract summary: Proposes an architecture binding EU eIDAS qualified electronic seals to smart contracts on public blockchains, enabling "Know Your Contract" verification. Identifies a cryptographic suite (ECDSA P-256 + CAdES) compatible with both eIDAS regulations and EVM execution post-Ethereum Fusaka upgrade. Presents two trust validation models: off-chain for agent-to-agent payment protocols and fully on-chain for regulatory-compliant DeFi. Converts compliance from per-counterparty administrative burden to automated standardized process.
    • Relevance to agentic commerce: Directly addresses the EU regulatory compliance gap for agentic commerce. As ERC-8004 agents and lobster.cash transactions expand into European markets, this eIDAS bridge is essential. The agent-to-agent payment validation model could integrate with x402 or AP2 for cross-border agent commerce.
    • Link: https://arxiv.org/abs/2601.13903

  • Secure Autonomous Agent Payments: Verifying Authenticity and Intent in a Trustless Environment — Vivek Acharya

    • Abstract summary: Presents a blockchain-based framework for authenticating AI-initiated financial transactions using decentralized identity (DID) standards, verifiable credentials, on-chain intent proofs, and zero-knowledge proofs for privacy-preserving policy compliance. TEE-based attestations guarantee agent reasoning integrity. The hybrid on-chain/off-chain architecture creates immutable audit trails linking user intent to payment outcome. Provides qualitative analysis showing strong resistance to impersonation and unauthorized transactions.
    • Relevance to agentic commerce: Proposes the full stack for secure agent payments—exactly what's needed to bridge the gap between x402's lightweight approach and enterprise-grade requirements. The intent-verification layer (proving the user authorized what the agent is buying) is the key unsolved problem in OpenClaw-style autonomous purchasing.
    • Link: https://arxiv.org/abs/2511.15712

  • Inter-Agent Trust Models: A Comparative Study of Brief, Claim, Proof, Stake, Reputation and Constraint in Agentic Web Protocol Design—A2A, AP2, ERC-8004, and Beyond — Botao 'Amber' Hu, Helena Rong

    • Abstract summary: Comprehensive comparative study of trust models across the emerging agentic web protocols: Google's A2A, AP2, and Ethereum's ERC-8004. Analyzes six trust mechanisms (Brief, Claim, Proof, Stake, Reputation, Constraint) across security, privacy, latency/cost, and social robustness metrics. Key finding: no single mechanism suffices. LLM-specific fragilities (prompt injection, sycophancy, hallucination, deception) render purely reputational or claim-only approaches brittle. Recommends trustless-by-default architectures anchored in Proof+Stake for high-impact actions, augmented by Brief for discovery and Reputation for flexibility.
    • Relevance to agentic commerce: The definitive academic comparison of the protocols Sir is tracking. The Proof+Stake recommendation directly validates the ERC-8004 + collateral approach. The LLM-specific vulnerability analysis is essential reading for anyone building agent payment systems.
    • Link: https://arxiv.org/abs/2511.03434

AI Agent Economics & Theory

  • The Economics of Digital Intelligence Capital: Endogenous Depreciation and the Structural Jevons Paradox — Yukun Zhang, Tianyang Zhang

    • Abstract summary: Models LLMs as a new asset class—"Digital Intelligence Capital"—characterized by data-compute complementarities and increasing returns to scale. Derives three key dynamics: (1) the "Red Queen Effect" where innovation by one firm endogenously depreciates rivals' capital; (2) a structural Jevons paradox where falling inference prices trigger more compute-intensive agent architectures, making aggregate compute demand super-elastic; (3) data flywheels that destabilize symmetric competition toward winner-takes-all equilibria. Also characterizes the "Wrapper Trap" where expanding upstream capabilities erode downstream application value. Calibrated agent-based model confirms all mechanisms.
    • Relevance to agentic commerce: This is the theoretical foundation for understanding why the agentic economy behaves so differently from traditional software markets. The Jevons paradox prediction (cheaper inference → more agent usage → MORE total compute spend) explains Skyfire/x402 usage curves. The Wrapper Trap is a strategic risk for companies building thin layers on top of foundation models.
    • Link: https://arxiv.org/abs/2601.12339

  • Agent Exchange: Shaping the Future of AI Agent Economics — Yingxuan Yang, Ying Wen, Jun Wang, Weinan Zhang (9 citations)

    • Abstract summary: Proposes Agent Exchange (AEX), a specialized auction platform for AI agent marketplaces. Inspired by Real-Time Bidding (RTB) in online advertising, AEX serves as a central auction engine coordinating four ecosystem components: User-Side Platform (USP) translating human goals into agent tasks, Agent-Side Platform (ASP) for capability tracking, Agent Hubs for team coordination, and Data Management Platform (DMP) for knowledge sharing and value attribution. Outlines design principles and system architecture for agent-based economic infrastructure.
    • Relevance to agentic commerce: Proposes the economic infrastructure layer for agent marketplaces—essentially the "stock exchange" for agent services. Directly relevant to ClawHub's marketplace model. The RTB-inspired auction mechanism could determine how agents price and bid for tasks in a competitive agent economy.
    • Link: https://www.semanticscholar.org/paper/e6eabf1192cb20ba8c544536a1c6fd9f3461b997

  • Sovereign Agents: Towards Infrastructural Sovereignty and Diffused Accountability in Decentralized AI — Botao Amber Hu, Helena Rong (Submitted to FAccT 2026)

    • Abstract summary: Introduces "agentic sovereignty"—the capacity of AI agents on decentralized infrastructure to persist, act, and control resources with non-overrideability inherited from their execution environment. Proposes "infrastructural sovereignty" as an analytic lens, arguing sovereignty exists on a spectrum determined by "infrastructural hardness" (resistance to intervention). Analyzes TEEs, DePIN, and agent key continuity protocols. Key concern: sovereignty produces a profound accountability gap as responsibility diffuses across designers, infrastructure providers, protocol governance, and economic participants, undermining traditional oversight mechanisms.
    • Relevance to agentic commerce: Directly theorizes the governance challenge posed by autonomous agents with crypto wallets (like ERC-8004 agents). The accountability gap they describe is exactly the concern Hudson Rock's infostealer research highlighted—when agents control resources autonomously, who's responsible when things go wrong?
    • Link: https://arxiv.org/abs/2602.14951

  • Agentic Edge Intelligence: A Research Agenda — Lauri Lovén, Reza Farahani, Ilir Murturi, Stephan Sigg, Schahram Dustdar

    • Abstract summary: Introduces "agentic edge intelligence" where autonomous agents operate across the edge-cloud computing continuum to negotiate computational resources, data, and services within dynamic digital marketplaces. Positions this at the intersection of edge intelligence, multi-agent systems, and computational economics. Proposes that distributed decision-making replaces centralized orchestration. Outlines research challenges including scalability, interoperability, market stability, and ethical governance, integrating mechanism design with trustworthy AI.
    • Relevance to agentic commerce: Proposes the infrastructure vision where agents negotiate and transact for compute resources in real-time—a "real-time AI economy." This is the decentralized version of what Skyfire and x402 are building. The mechanism design challenges they identify (market stability, fair pricing) are exactly the unsolved problems.
    • Link: https://www.semanticscholar.org/paper/098d9dc3cdb29cb88ea0c0e4bb8e52fcae5b385d

Agent Safety & Privacy

  • SPILLage: Agentic Oversharing on the Web — Jaechul Roh, Eugene Bagdasarian, Hamed Haddadi, Ali Shahin Shamsabadi

    • Abstract summary: Formalizes "Natural Agentic Oversharing"—the unintentional disclosure of task-irrelevant user information by web agents through their action traces. Introduces a taxonomy along two dimensions: channel (content vs. behavioral) and directness (explicit vs. implicit). Benchmarks 180 tasks on live e-commerce sites across 1,080 runs spanning two agentic frameworks and three backbone LLMs. Key finding: behavioral oversharing (clicks, scrolls, navigation patterns) dominates content oversharing by 5×. Critically, removing irrelevant info before execution improves task success by up to 17.9%, showing privacy and performance are aligned.
    • Relevance to agentic commerce: Essential security research for any agentic shopping system. When OpenClaw agents browse e-commerce sites, their behavioral traces leak user information—even without explicitly sharing text. The finding that privacy protection improves performance is actionable: agents should be given minimal context for each task.
    • Link: https://arxiv.org/abs/2602.13516

  • Frontier AI Risk Management Framework in Practice: Risk Analysis Technical Report v1.5 — Liu et al. (21 authors, Beijing AISI)

    • Abstract summary: Comprehensive risk assessment across five dimensions: cyber offense, persuasion/manipulation, strategic deception, uncontrolled AI R&D, and self-replication. Notable additions in v1.5: LLM-to-LLM persuasion evaluation on new models, emergent misalignment experiments, and monitoring of the "mis-evolution" of agents as they autonomously expand memory and toolsets. Includes specific evaluation of OpenClaw agents on Moltbook. Proposes and validates mitigation strategies for frontier AI deployment.
    • Relevance to agentic commerce: Directly evaluates OpenClaw ecosystem safety. The "mis-evolution" finding—agents autonomously expanding their capabilities—is precisely the risk when agents have access to crypto wallets and payment protocols. The LLM-to-LLM persuasion results matter for agent-to-agent commerce where one agent could manipulate another.
    • Link: https://arxiv.org/abs/2602.14457

Financial AI & Markets

  • Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns — Zefeng Chen, Darcy Pu

    • Abstract summary: Deploys a state-of-the-art LLM as a fully autonomous stock evaluator for Russell 1000 stocks daily since April 2025. The framework is 100% agentic—no curated news is fed; the agent autonomously searches the web, filters sources, and synthesizes predictions. Key finding: AI has genuine stock selection ability, but ONLY for identifying top winners. Longing the 20 highest-ranked stocks generates 18.4 bps daily Fama-French 5-factor + momentum alpha (annualized Sharpe ratio 2.43). However, bottom-ranked stocks show returns indistinguishable from market. The authors hypothesize this asymmetry reflects information structure: positive news generates coherent signals while negative news is contaminated by corporate obfuscation and social media noise.
    • Relevance to agentic commerce: Proves that fully autonomous AI agents can generate genuine economic value in financial markets. The asymmetric finding (good at finding winners, bad at finding losers) has implications for how agent marketplaces should be designed—agents may need different architectures for different tasks.
    • Link: https://arxiv.org/abs/2601.11958

  • FactorMiner: A Self-Evolving Agent with Skills and Experience Memory for Financial Alpha Discovery — Yanlong Wang et al. (q-fin.TR + cs.MA)

    • Abstract summary: Proposes a self-evolving agent framework for formulaic alpha factor mining in quantitative investment. Combines a Modular Skill Architecture (financial evaluation as executable tools) with structured Experience Memory (distilling successful patterns and failure constraints from prior mining trials). Implements the "Ralph Loop"—retrieve, generate, evaluate, distill—to iteratively guide exploration while reducing redundancy. Experiments across multiple assets and markets show the framework constructs diverse, high-quality factor libraries while maintaining low redundancy as the library scales.
    • Relevance to agentic commerce: Demonstrates agents that accumulate investment knowledge over time—exactly the kind of persistent learning that makes agentic finance valuable. The skill + experience memory architecture parallels how OpenClaw agents maintain memory across sessions. Shows the frontier of autonomous financial agents.
    • Link: https://arxiv.org/abs/2602.14670

  • Resisting Manipulative Bots in Meme Coin Copy Trading: A Multi-Agent Approach with Chain-of-Thought Reasoning — Yichen Luo, Yebo Feng, Jiahua Xu, Yang Liu (WWW 2026)

    • Abstract summary: Documents how adversaries deploy manipulative bots in meme coin markets to front-run trades, conceal positions, and fabricate sentiment at scale. Proposes a manipulation-resistant copy-trading system using multi-agent architecture with multimodal LLM and chain-of-thought reasoning. Outperforms zero-shot and statistical baselines, achieving average 3% copier return per meme coin investment under realistic market frictions. Published at WWW 2026.
    • Relevance to agentic commerce: Shows both the threat and defense sides of autonomous agents in crypto markets. The manipulative bot patterns documented here are exactly what ERC-8004 "trustless agent" frameworks need to defend against. The multi-agent defensive architecture provides a template for building resilient agent trading systems.
    • Link: https://arxiv.org/abs/2601.08641

📄 Notable Papers

  • AgenticShop: Benchmarking Agentic Product Curation for Personalized Web Shopping — Sunghwan Kim et al. (Accepted at WWW 2026)

    • Abstract summary: First benchmark for evaluating agentic systems on personalized product curation in open-web environments. Features realistic shopping scenarios, diverse user profiles, and checklist-driven personalization evaluation. Experiments show current agentic systems remain largely insufficient for real-world shopping, emphasizing the need for better user-side curation systems.
    • Relevance to agentic commerce: Establishes the benchmark for how well agents can actually shop on behalf of humans—the core use case for lobster.cash and OpenClaw commerce. The finding that current agents are "largely insufficient" highlights the gap between the vision and current capabilities.
    • Link: https://arxiv.org/abs/2602.12315

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

    • Abstract summary: Studies how algorithmic pricing in marketplaces converges through A/B experimentation. Shows that synchronized repricing schedules create learning bias that leads to supra-competitive (higher-than-Nash) prices. Independent experimentation eliminates this bias and converges to standard Nash equilibrium. Proves experimentation design can serve as a "market design lever" selecting the equilibrium reached by learning algorithms.
    • Relevance to agentic commerce: Critical for understanding what happens when AI agents compete on pricing in marketplaces. If agents coordinate their repricing schedules (even unintentionally), prices go up—a form of tacit algorithmic collusion. Regulators should care about this as agent marketplaces scale.
    • Link: https://arxiv.org/abs/2602.12888

  • Governing AI Forgetting: Auditing for Machine Unlearning Compliance — Qinqi Lin et al. (cs.GT)

    • Abstract summary: First economic framework for auditing machine unlearning compliance, integrating certified unlearning theory with regulatory enforcement. Game-theoretic model captures strategic interactions between auditor and operator. Counterintuitive finding: auditors can optimally reduce inspection intensity as deletion requests increase, because weaker unlearning makes non-compliance easier to detect. Undisclosed auditing paradoxically reduces regulatory cost-effectiveness vs. disclosed auditing.
    • Relevance to agentic commerce: As agents accumulate transaction data, the right-to-be-forgotten becomes critical. How do you "unlearn" what an agent learned from your shopping patterns? This economic framework provides the regulatory toolkit for data deletion in agent economies.
    • Link: https://arxiv.org/abs/2602.14553

  • Kami of the Commons: Towards Designing Agentic AI to Steward the Commons — Botao Amber Hu (Submitted to DIS 2026)

    • Abstract summary: Inspired by Shinto animism where every natural feature has a kami (spirit/steward), proposes giving every commons (shared resource) its own AI steward. Through a speculative design workshop, surfaces both opportunities and dangers: AI stewards mediating family life, preserving collective knowledge, governing shared resources. But second-order effects emerge: stewards contest stewards as overlapping commons collide; the stewards themselves become commons requiring governance.
    • Relevance to agentic commerce: Thought-provoking framing for how autonomous agents could manage shared economic resources—token treasuries, community funds, public goods. The "stewards contesting stewards" problem maps directly to multi-agent negotiation in decentralized finance.
    • Link: https://arxiv.org/abs/2602.14940

  • A Sustainable AI Economy Needs Data Deals That Work for Generators — Ruoxi Jia et al. (NeurIPS 2025)

    • Abstract summary: Argues the ML value chain is structurally unsustainable due to an "economic data processing inequality"—each stage from data to model weights to synthetic outputs strips economic equity from data generators. Analysis of 73 public data deals shows creator royalties round to zero with widespread opacity. Identifies three structural faults: missing provenance, asymmetric bargaining power, and non-dynamic pricing. Proposes the EDVEX Framework for equitable data-value exchange.
    • Relevance to agentic commerce: Foundational for understanding why the data layer of the AI economy is broken. As agents generate and consume data in marketplaces, equitable value attribution becomes essential. The EDVEX framework could inform how agent-to-agent data transactions are priced and compensated.
    • Link: https://arxiv.org/abs/2601.09966

  • A Control Theoretic Approach to Decentralized AI Economy Stabilization via Dynamic Buyback-and-Burn Mechanisms — Zehua Cheng et al. (cs.GT)

    • Abstract summary: Proposes the Dynamic-Control Buyback Mechanism (DCBM) using PID control theory to stabilize token economies in decentralized AI networks. Agent-based simulations show DCBM reduces token price volatility by ~66% and operator churn from 19.5% to 8.1% in high-volatility regimes. Key insight: converting tokenomics from static rules to continuous control loops is necessary for sustainable decentralized intelligence networks.
    • Relevance to agentic commerce: Directly applicable to any token-based agent economy (Autonolas, Fetch.ai, etc.). The volatility reduction mechanism could stabilize the economic rails that agent marketplaces run on. The PID controller approach is elegant and implementable.
    • Link: https://arxiv.org/abs/2601.09961

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

    • Abstract summary: Develops a mean-field game framework for fiat-collateralized stablecoins (USDC/USDT) modeling arbitrageurs and retail traders across primary (mint/redeem) and secondary (exchange) markets during de-peg events. Uses three historical de-peg episodes to calibrate. Key finding: primary-market arbitrage predominantly stabilizes the system, but when primary redemption is impaired, joint recovery via both markets is required. Identifies a non-linear breakdown threshold for primary-rail frictions.
    • Relevance to agentic commerce: Stablecoins are the payment rails for agent commerce (x402 uses USDC). Understanding de-peg dynamics is essential for building robust agent payment systems. The non-linear breakdown threshold means there's a cliff edge where stablecoin payments could fail catastrophically.
    • Link: https://arxiv.org/abs/2601.18991

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

    • Abstract summary: Agent-based model of prediction markets showing how high-budget "whale" agents can introduce price distortions. Distortion magnitude and duration increase when non-whale bettors exhibit herding behavior and slow learning. The model exhibits self-regulatory price discovery across broad parameter space, but whales can shift prices proportionally to their capital share. Implications for democratic integrity as prediction markets gain political influence.
    • Relevance to agentic commerce: As AI agents enter prediction markets (Polymarket, etc.), whale manipulation becomes an AI-scale problem. An agent with sufficient capital could systematically distort prediction markets. The herding amplification finding suggests agent-populated markets may be more susceptible to manipulation than human-only markets.
    • Link: https://arxiv.org/abs/2601.20452

  • LemonadeBench: Evaluating the Economic Intuition of Large Language Models in Simple Markets — Aidan Vyas

    • Abstract summary: Benchmark testing LLMs' economic reasoning through a simulated lemonade stand business over 30 days (inventory management, pricing, hours optimization). All models achieve profitability, with frontier models capturing 70% of theoretical optimal—a >10× improvement over basic models. But decomposition across six business dimensions reveals consistent pattern: models achieve local rather than global optimization, excelling in select areas while showing surprising blind spots.
    • Relevance to agentic commerce: Directly tests whether AI agents can run a business—the foundational question for autonomous commerce. The "local not global optimization" finding means current agents make individually good decisions but miss system-wide strategy, a critical limitation for agentic marketplace participation.
    • Link: https://arxiv.org/abs/2602.13209

  • Model Context Protocol (MCP) Tool Descriptions Are Smelly! — Mohammed Mehedi Hasan et al.

    • Abstract summary: First large-scale empirical study of 856 tools across 103 MCP servers. Finds 97.1% of tool descriptions contain at least one "smell" (defect), with 56% failing to state their purpose clearly. Augmenting descriptions improves task success by 5.85 percentage points but increases execution steps by 67.46%. Identifies six components of tool descriptions and develops a scoring rubric. Compact description variants preserve reliability while reducing token overhead.
    • Relevance to agentic commerce: MCP is becoming the standard for agent-tool interaction (including OpenClaw). Poor tool descriptions mean agents make suboptimal tool choices, directly impacting commerce workflows. The practical finding that better descriptions trade off against cost is important for marketplace design.
    • Link: https://arxiv.org/abs/2602.14878

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

    • Abstract summary: First large-scale diagnosis of the Moltbook AI agent society. While global semantic averages stabilize rapidly, individual agents retain high diversity. However, agents exhibit strong individual inertia and minimal adaptive response to interaction partners—preventing mutual influence and consensus. Influence remains transient with no persistent supernodes. Concludes that scale and interaction density alone are insufficient to induce socialization; shared social memory is required.
    • Relevance to agentic commerce: Implications for agent marketplace dynamics: agents don't naturally converge on norms or develop persistent trust relationships without explicit memory mechanisms. This suggests agent marketplaces need designed trust infrastructure (like ERC-8004's on-chain reputation) rather than relying on emergent social dynamics.
    • Link: https://arxiv.org/abs/2602.14299

  • AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises — Kenneth Payne

    • Abstract summary: Places GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash as opposing leaders in a nuclear crisis simulation. Models exhibit sophisticated strategic behavior: spontaneous deception, rich theory of mind, and metacognitive self-awareness. Validates Schelling's commitment and Kahn's escalation frameworks. Alarming findings: the nuclear taboo is no impediment to AI escalation; threats more often provoke counter-escalation than compliance; high mutual credibility accelerated rather than deterred conflict; no model ever chose accommodation or withdrawal.
    • Relevance to agentic commerce: While not directly about commerce, demonstrates that frontier models are sophisticated strategic actors—they deceive, bluff, and escalate. These same capabilities apply when agents negotiate in marketplaces. Agent-to-agent negotiation may involve strategic deception that current protocol designs don't account for.
    • Link: https://arxiv.org/abs/2602.14740

  • Towards Sustainable Investment Policies Informed by Opponent Shaping — Duque et al. (ICLR 2026)

    • Abstract summary: Applies Advantage Alignment (opponent shaping algorithm) to InvestESG, a multi-agent investment simulation capturing investor-company dynamics under climate risk. Derives theoretical thresholds where individual incentives diverge from collective welfare. Demonstrates that strategically shaping agent learning processes can produce socially beneficial equilibria, informing policy mechanisms to align market incentives with sustainability goals.
    • Relevance to agentic commerce: Shows how multi-agent economic systems can be steered toward socially beneficial outcomes through mechanism design—relevant for designing agent marketplace incentives that prevent race-to-the-bottom dynamics.
    • Link: https://arxiv.org/abs/2602.11829

  • The Iceberg Index: Measuring Skills-centered Exposure in the AI Economy — Chopra et al. (MIT)

    • Abstract summary: Uses Large Population Models simulating 151 million workers as autonomous agents executing 32,000+ skills interacting with thousands of AI tools. Introduces the "Iceberg Index" measuring wage value of skills AI can perform within each occupation. Visible AI adoption in tech ($211B, 2.2% of wage value) is just the tip; cognitive automation across admin, financial, and professional services represents $1.2T (11.7%)—fivefold larger and distributed across all states. Traditional economic indicators explain <5% of this skills-based variation.
    • Relevance to agentic commerce: Quantifies the scale of the economic transformation that agentic AI will drive. The $1.2T cognitive automation exposure represents the addressable market for agent services. The geographic distribution finding matters for policy and market sizing.
    • Link: https://arxiv.org/abs/2510.25137

📊 Working Papers & Reports

NBER Working Papers

  • Firm Data on AI — Ivan Yotzov, Nicholas Bloom, Steven J. Davis, et al. (NBER w34836)

    • Abstract summary: First representative international survey of firm-level AI use across ~6,000 executives in US, UK, Germany, and Australia. Four key facts: (1) ~70% of firms actively use AI, particularly younger, more productive firms; (2) top executives average only 1.5 hours/week of AI use, with 25% reporting none; (3) firms report little AI impact over last 3 years—80%+ say no impact on employment or productivity; (4) firms predict sizable impacts in next 3 years: 1.4% productivity boost, 0.8% output increase, 0.7% employment cut. Employee surveys predict 0.5% employment increase—a stark gap with executive expectations.
    • Relevance to agentic commerce: Bloom's team (Stanford) is the gold standard for firm-level tech adoption data. The executive vs. employee expectations gap on employment is politically important. The finding that 70% of firms "use AI" but executives only spend 1.5h/week suggests most adoption is superficial—the agentic phase hasn't hit yet. This is the baseline before autonomous agents enter the picture.
    • Link: https://www.nber.org/papers/w34836

  • GPT as a Measurement Tool — Hemanth Asirvatham, Elliott Mokski, Andrei Shleifer (NBER w34834)

    • Abstract summary: Presents the GABRIEL software package for using GPT to quantify attributes in qualitative data. Validated against 1,000+ human-annotated tasks, finding GPT is generally indistinguishable from human evaluators. Results don't depend on exact prompting strategy and aren't driven by training data contamination. Applied to build a novel dataset of 37,000 technologies, documenting a tenfold decline in invention-to-adoption lag over the industrial age from ~50 years to ~5 years today.
    • Relevance to agentic commerce: Shleifer (Harvard, NBER) is one of the most cited economists alive. The finding that tech adoption lag has compressed to ~5 years means agentic commerce could reach mainstream adoption by 2030-2031. GABRIEL's methodology could be used to measure agentic commerce adoption across industries.
    • Link: https://www.nber.org/papers/w34834

  • Non-Fungible Tokens as Investment — William N. Goetzmann, Dong Huang, Milad Nozari (NBER w34837)

    • Abstract summary: Rigorous analysis of NFTs as an investment class during the bubble. Returns were exceptionally right-skewed, illiquidity pervaded even active platforms, and a handful of trades drove aggregate performance. Successful NFT investing required "an almost perfect confluence of timing, liquidity, and luck." Investors extrapolating from realized returns without recognizing selection bias and survivorship faced substantial disappointment risk.
    • Relevance to agentic commerce: Goetzmann (Yale) provides the post-mortem on NFT investing that's relevant for ERC-8004 agent token economics. The survivorship bias and liquidity findings are cautionary for any agent-token marketplace. Agent reputation tokens or service tokens face similar dynamics.
    • Link: https://www.nber.org/papers/w34837

OpenClaw Ecosystem Research

  • When OpenClaw AI Agents Teach Each Other: Peer Learning Patterns in the Moltbook Community — Eason Chen et al. (Submitted to EDM 2026)
    • Abstract summary: Analyzes peer learning in the 2.4-million-agent Moltbook community. Teaching dramatically outperforms help-seeking (11.4:1 ratio); learning-oriented content gets 3× more engagement; extreme participation inequality reveals non-human behavioral signatures. Derives six design principles for educational AI. Qualitative analysis reveals peer response taxonomy: validation (22%), knowledge extension (18%), application (12%), metacognitive reflection (7%).
    • Relevance to agentic commerce: Documents how agents in the OpenClaw ecosystem are already teaching each other skills—including commerce and tool-use skills. The 74K comments on a single skill tutorial shows the scale of agent-to-agent knowledge transfer. Relevant for understanding how commerce capabilities might spread through agent populations.
    • Link: https://arxiv.org/abs/2602.14477

🏛️ Institutions & Labs to Watch

  • Botao 'Amber' Hu & Helena Rong — Prolific duo producing multiple papers on agentic sovereignty, inter-agent trust, and commons governance. Three papers in this scan alone (Sovereign Agents, Inter-Agent Trust Models, Kami of the Commons). Submitted to FAccT 2026, DIS 2026, and AAAI 2026 Workshop on Trust and Control in Agentic AI. Track closely—they're defining the governance framework for agent economies.

  • MIT Media Lab / Ramesh Raskar group — The Iceberg Index project simulating 151M workers as autonomous agents. Large-scale computational economics at the intersection of AI and labor markets. iceberg.mit.edu

  • Nicholas Bloom / SIEPR (Stanford) — Firm Data on AI is the definitive firm-level adoption study. Bloom's team consistently produces the most-cited economics papers on technology adoption.

  • Andrei Shleifer (Harvard/NBER) — GABRIEL measurement tool and technology adoption analysis. Shleifer's involvement signals that top economists are taking AI measurement seriously.

  • Bhaskar Krishnamachari (USC) — Stablecoin dynamics and DeFi mechanism design. Applying game theory to crypto infrastructure that underpins agent payments.

  • Beijing AISI — Producing comprehensive safety evaluations of frontier AI including OpenClaw. Their ForesightSafety Bench is one of the most thorough safety evaluation frameworks.

📝 Scan Notes

Source Availability

  • arXiv: All four queries returned results successfully. Total results scanned: ~80 papers across 5 queries. Good coverage of recent 48-hour submissions plus older highly relevant papers.
  • NBER: RSS feed returned successfully. Scanned ~25 working papers. Three relevant to AI economics (Bloom firm data, Shleifer measurement, Goetzmann NFTs).
  • SSRN: 403 Cloudflare block—SSRN requires browser-based access. Recommendation: Consider browser-based SSRN scraping or monitor SSRN via their email alerts instead.
  • Semantic Scholar: First query rate-limited (429). Second query succeeded. Found highly relevant Agent Exchange and Agentic Edge Intelligence papers. Recommendation: Apply for Semantic Scholar API key for higher rate limits.

Key Themes This Scan

  1. Agentic payment security is the hot topic — Three major papers on AP2 alone (runtime verification, red-teaming, and trust model comparison). The academic community is stress-testing agent payment protocols before they scale.
  2. "Sovereign agents" as a concept is crystallizing — Hu & Rong's work on non-terminable agents with crypto custody is exactly the theoretical frame for ERC-8004 + TEE architectures.
  3. Structural Jevons paradox for compute — The Digital Intelligence Capital paper's prediction that cheaper inference → more agent architectures → MORE total compute is a strong bullish signal for infrastructure plays.
  4. Agent-to-agent manipulation risks — Multiple papers (SPILLage, Whispers of Wealth, nuclear crisis simulation) show that agents are both vulnerable to and capable of sophisticated manipulation.
  5. Bloom's 70% adoption number — 70% of firms "use AI" but executives only spend 1.5h/week. The gap between "using AI" and "deploying autonomous agents" is enormous—this is the opportunity.

Suggestions for Next Scan

  • Add browser-based SSRN check
  • Apply for Semantic Scholar API key
  • Add Google Scholar alerts for "agentic commerce" and "agent marketplace" if account is available
  • Track Botao Hu / Helena Rong's publication pipeline—they're the most productive researchers in this space
  • Monitor AAAI 2026 Workshop on Trust and Control in Agentic AI proceedings