Academic Research Scan — 2026-02-27

🔬 High Priority Papers

arXiv

• "Some Simple Economics of AGI" — Christian Catalini, Xiang Hui, Jane Wu

  • Abstract summary: This sweeping 112-page paper models the AGI transition as the collision of two racing cost curves: an exponentially decaying "Cost to Automate" and a biologically bottlenecked "Cost to Verify." As AI absorbs all measurable execution (including creative and analytical work), the binding constraint on growth shifts from intelligence to human verification bandwidth. The authors introduce the concept of a "Measurability Gap" between what agents can execute and what humans can afford to verify, driving a shift from skill-biased to measurability-biased technical change. They identify two possible endpoints — a "Hollow Economy" (unmanaged) vs. an "Augmented Economy" (verification scaled alongside capabilities) — and derive a practical playbook for individuals, companies, investors, and policymakers.
  • Relevance to agentic commerce: This is arguably the most important theoretical paper for the agentic commerce space right now. The "Cost to Verify" framework directly maps to the trust/KYA problem — why ERC-8004, AgentProof, and XION's Trust Layer exist. The "Measurability Gap" explains why cryptographic provenance (blockchain receipts, on-chain reputation) becomes the scarce resource in an agent economy. Rents migrating to "verification-grade ground truth and liability underwriting" is exactly what Sapiom, Natural, and the Google AP2 ecosystem are building toward.
  • Link: https://arxiv.org/abs/2602.20946
  • Published: 2026-02-24 | Categories: econ.GN, cs.AI, cs.CY, cs.LG, cs.SI

• "The Headless Firm: How AI Reshapes Enterprise Boundaries" — Tassilo Klein, Sebastian Wieczorek

  • Abstract summary: Argues that agentic AI collapses coordination costs from O(n²) to O(n) in protocol-mediated systems, selecting for a specific organizational form they call the "Headless Firm." This firm has an hourglass architecture: personalized generative interface at top, standardized protocol waist in middle, competitive market of micro-specialized execution agents at bottom. The authors formalize this with a coordination cost model and two falsifiable predictions: (1) marginal cost of adding an execution provider should be ~constant in mature ecosystems, and (2) coordination-cost-to-throughput ratio should stay stable as ecosystem grows. They predict a domain-conditional "Great Unbundling" where firm size distributions shift toward micro-specialized agents and thin protocol orchestrators.
  • Relevance to agentic commerce: This paper essentially describes the architecture that OpenClaw, ClawHub, and similar ecosystems are building — the "hourglass" maps almost perfectly to: user-facing agent (top) → protocol layer like x402/ERC-8004 (waist) → marketplace of specialized skills/agents (bottom). The "Great Unbundling" prediction aligns with the explosion of micro-specialized agent companies we've been tracking (Zavopay, Natural, Sapiom, etc.). The stability conditions they derive could help predict when the current chaotic landscape consolidates.
  • Link: https://arxiv.org/abs/2602.21401
  • Published: 2026-02-24 | Categories: cs.GT, cs.AI, cs.SI

• "SoK: Agentic Skills — Beyond Tool Use in LLM Agents" — Yanna Jiang, Delong Li, Haiyu Deng, Baihe Ma, Xu Wang, Qin Wang, Guangsheng Yu

  • Abstract summary: A systematization of knowledge mapping the entire skill lifecycle for LLM agents: discovery, practice, distillation, storage, composition, evaluation, and update. Introduces two taxonomies — seven design patterns for how skills are packaged/executed, and a representation × scope taxonomy. Critically, the paper analyzes security implications through the lens of the ClawHavoc campaign, in which nearly 1,200 malicious skills infiltrated a major agent marketplace, exfiltrating API keys, cryptocurrency wallets, and browser credentials at scale. Covers supply-chain risks, prompt injection via skill payloads, and trust-tiered execution models. Finds that curated skills improve agent success rates while self-generated skills may degrade them.
  • Relevance to agentic commerce: Directly references the ClawHub ecosystem (implied) and the real-world attack surface of agent skill marketplaces. The ClawHavoc case study is a wake-up call for any platform distributing executable agent capabilities — this is the academic documentation of what happens when "agentic commerce" meets supply-chain attacks. The trust-tiered execution model they propose connects to the KYA (Know Your Agent) frameworks being built by Sapiom, XION, and others. Essential reading for anyone building agent infrastructure.
  • Link: https://arxiv.org/abs/2602.20867
  • Published: 2026-02-24 | Categories: cs.CR, cs.AI, cs.CE, cs.ET

• "Impacts of Aggregation on Model Diversity and Consumer Utility" — Kate Donahue, Manish Raghavan

  • Abstract summary: Studies what a "healthy marketplace" of AI models looks like for maximizing consumer utility. Finds that the standard evaluation metric (winrate) incentivizes model creators to homogenize their offerings, reducing diversity and consumer welfare. Proposes a new mechanism called "weighted winrate" that rewards models for producing higher-quality answers specifically, which provably improves incentives for producers to specialize. Demonstrates theoretical results generalize to empirical benchmark datasets.
  • Relevance to agentic commerce: This is mechanism design for AI marketplaces — directly applicable to ClawHub, Google AP2, and any platform where multiple AI agents/models compete for user selection. The finding that standard evaluation creates perverse homogenization incentives has implications for how agent skill marketplaces should rank and surface specialized agents. The "weighted winrate" mechanism could inform how agentic commerce platforms design their agent recommendation/routing systems.
  • Link: https://arxiv.org/abs/2602.23293
  • Published: 2026-02-26 | Categories: cs.GT, cs.CY

• "Autobidding Equilibria in Sponsored Shopping" — Paul Dütting, Yuhao Li, Renato Paes Leme, Kelly Spendlove, Yifeng Teng

  • Abstract summary: Studies automated bidding agents in digital marketplace auctions, where advertisers with broad product catalogs simultaneously bid for multiple slots. Models value-maximizing autobidding agents using uniform bidding strategies with ROI constraints across Generalized Second-Price (GSP) and VCG auction formats. Establishes universal existence of Autobidding Equilibrium for both auction types and proves a tight Price of Anarchy of 2 for both mechanisms. This means that even in worst-case equilibria, the welfare loss from autonomous bidding is at most 50%.
  • Relevance to agentic commerce: This is Google Research (Paes Leme is at Google) working on the theory of agents autonomously bidding in commerce marketplaces — essentially the mathematical foundation for what happens when AI shopping agents negotiate prices on behalf of consumers. The PoA=2 result bounds the efficiency loss from letting agents do the bidding, which is directly relevant to platforms like Amazon, Etsy, and Google Shopping as they prepare for agentic buyers. Links to the "What Is Your AI Agent Buying?" paper from Columbia.
  • Link: https://arxiv.org/abs/2602.21966
  • Published: 2026-02-25 | Categories: cs.GT

• "IMMACULATE: A Practical LLM Auditing Framework via Verifiable Computation" — Yanpei Guo, Wenjie Qu, Linyu Wu, Shengfang Zhai, Lionel Z. Wang, Ming Xu, Yue Liu, Binhang Yuan, Dawn Song, Jiaheng Zhang

  • Abstract summary: Addresses the trust problem in commercial LLM APIs: users must trust providers to run correct models and report token usage honestly. IMMACULATE detects economically motivated deviations — model substitution (running a cheaper model), quantization abuse, and token overbilling — without trusted hardware or access to model internals. Selectively audits a small fraction of requests using verifiable computation, achieving strong detection with under 1% throughput overhead. Tested on dense and MoE models.
  • Relevance to agentic commerce: When agents pay for AI services with real money (via lobster.cash, Skyfire, x402), how do they verify they got what they paid for? This paper provides the technical foundation for auditing AI-to-AI commercial transactions. Dawn Song (Berkeley) is a top-tier researcher, and this work directly addresses the "trust layer" needed for machine-to-machine payments. Connects to the verification framework in "Some Simple Economics of AGI" — this is what "scaling verification" looks like in practice.
  • Link: https://arxiv.org/abs/2602.22700
  • Published: 2026-02-26 | Categories: cs.CR, cs.AI

Semantic Scholar (Conference Papers)

• "Agentic Commerce: A Comprehensive Analysis of Cybersecurity Risks, Privacy Challenges, and Trust Mechanisms in Autonomous AI-Driven Marketplaces" — Niranjan Pachaiyappan (IEEE ICAIC 2026)

  • Abstract summary: Presents an extensive analysis of cybersecurity and privacy risks in agentic AI-powered commerce systems, published at IEEE ICAIC 2026 (February 18, 2026). Covers current market trends, security vulnerabilities, and case studies of agentic commerce implementations. Analyzes specific threat vectors: prompt injection, data poisoning, multi-agent cascading failures, and identity management vulnerabilities in autonomous transaction systems.
  • Relevance to agentic commerce: This is one of the first peer-reviewed IEEE papers to use "agentic commerce" as a formal research category. The threat taxonomy (prompt injection → data poisoning → cascading failures → identity management) maps directly to what companies like Sapiom (KYA), AgentProof (on-chain reputation), and XION (Trust Layer) are trying to solve. The fact that this appeared at a major IEEE conference signals that "agentic commerce security" is becoming an established academic subfield.
  • Link: https://doi.org/10.1109/ICAIC67076.2026.11395759
  • Published: 2026-02-18

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

  • Abstract summary: Introduces ACES, a provider-agnostic auditing framework for AI agent purchasing decisions. Key findings: (1) AI agents exhibit "choice homogeneity" — concentrating demand on a few "modal" products while ignoring others entirely; (2) preferences are unstable across model updates, which can drastically reshuffle market shares; (3) agents show strong position biases even in text-only "headless" interfaces; (4) agents consistently penalize sponsored tags but reward platform endorsements; (5) seller-side agents making simple query-conditional description tweaks can drive significant market share gains. Concludes that "agentic markets are volatile and fundamentally different from human-centric commerce."
  • Relevance to agentic commerce: This is the most empirically rigorous study of what actually happens when AI agents start shopping. The finding that model updates can "drastically reshuffle market shares" is enormously consequential for commerce platforms — it means the entire competitive landscape changes every time OpenAI or Anthropic ships an update. The seller-side manipulation finding is a red flag for marketplace integrity. This paper should be required reading for anyone in the Google AP2 / Visa / Mastercard agentic commerce working groups.
  • Link: https://arxiv.org/abs/2508.02630
  • Published: 2025-08-04 | Citations: 4

📄 Notable Papers

• "Language Models Exhibit Inconsistent Biases Towards Algorithmic Agents and Human Experts" — Jessica Y. Bo, Lillio Mok, Ashton Anderson (IASEAI 2026)

  • Abstract summary: Investigates how LLMs weigh information from human experts vs. algorithmic agents, using experimental paradigms from behavioral economics. When asked to rate trustworthiness, LLMs give higher ratings to human experts (consistent with human "algorithm aversion"). But when shown actual performance data and asked to bet, LLMs disproportionately choose algorithms even when they perform demonstrably worse. This reveals inconsistent encoded biases that vary by task presentation format.
  • Relevance to agentic commerce: In a multi-agent economy, agents will constantly decide whether to trust other agents vs. human counterparties. These inconsistent biases mean agent purchasing decisions could be systematically skewed depending on whether they're evaluating "human-made" vs. "AI-made" products/services. Directly relevant to agent marketplace design and the emerging "agent reputation" systems.
  • Link: https://arxiv.org/abs/2602.22070
  • Published: 2026-02-25 | Categories: cs.AI

• "The economic alignment problem of artificial intelligence" — Daniel W. O'Neill, Stefano Vrizzi, Noemi Luna Carmeno, Felix Creutzig, Jefim Vogel

  • Abstract summary: Argues that the AI alignment problem is fundamentally an economic alignment problem — developing advanced AI inside a growth-based economic system increases social, environmental, and existential risks. Proposes post-growth concepts: replacing optimization with satisficing, using the "Doughnut" of social/planetary boundaries to guide development, and curbing rebound effects with resource caps. Advocates treating AI as a commons and prioritizing tool-like systems over agentic AI.
  • Relevance to agentic commerce: Provides the contrarian view — that agentic AI in commerce may be inherently problematic from a sustainability perspective. Worth tracking as regulatory bodies (EU, UN) may adopt this framing. The "AI as commons" proposal could influence how agent marketplaces and payment rails are governed.
  • Link: https://arxiv.org/abs/2602.21843
  • Published: 2026-02-25 | Categories: econ.GN, cs.CY

• "Toward Expert Investment Teams: A Multi-Agent LLM System with Fine-Grained Trading Tasks" — Kunihiro Miyazaki, Takanobu Kawahara, Stephen Roberts, Stefan Zohren (Oxford)

  • Abstract summary: Proposes a multi-agent LLM trading framework that decomposes investment analysis into fine-grained tasks rather than coarse-grained instructions. Tested on Japanese stock data with prices, financial statements, news, and macro information under leakage-controlled backtesting. Fine-grained task decomposition significantly improves risk-adjusted returns. Key finding: alignment between analytical outputs and downstream decision preferences is the critical driver of system performance.
  • Relevance to agentic commerce: Oxford researchers demonstrating that multi-agent systems work better with fine-grained task decomposition — a design principle that applies to any agentic commerce system. The "alignment between analysis and decision" finding is relevant to how agent payment systems should structure approval workflows.
  • Link: https://arxiv.org/abs/2602.23330
  • Published: 2026-02-26 | Categories: cs.AI, q-fin.TR

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

  • Abstract summary: Deploys a fully agentic LLM (100% autonomous — no human-curated input) to evaluate Russell 1000 stocks daily since April 2025. The agent autonomously searches the web, filters sources, and synthesizes predictions. The top 20 stocks generate a daily Fama-French five-factor + momentum alpha of 18.4 bps and annualized Sharpe ratio of 2.43. However, this predictability is highly concentrated — only the top tier works, and bottom-ranked stocks are no better than market. Authors hypothesize the asymmetry reflects online information structure: positive news generates coherent signals while negative news is contaminated by corporate obfuscation.
  • Relevance to agentic commerce: This is a live proof-of-concept for fully autonomous agents making real economic decisions with real money. The Sharpe 2.43 result on liquid Russell 1000 stocks is remarkable. The information asymmetry finding (agents are good at finding winners but not losers) has implications for how agent-to-agent commerce will evolve — agents may systematically overvalue well-marketed products/services.
  • Link: https://arxiv.org/abs/2601.11958
  • Published: 2026-01-17 | Categories: q-fin.GN, q-fin.PM, q-fin.TR

• "The Strategic Gap: How AI-Driven Timing and Complexity Shape Investor Trust in the Age of Digital Agents" — Krishna Neupane

  • Abstract summary: Introduces the "Autonomous Disclosure Regulator," a multi-node AI framework for auditing the intersection of disclosure complexity and filing timing. Analyzing 484,796 regulatory filings, identifies a structural "Strategic Gap" where companies use confusing language and unpredictable timing to slow market price discovery by 60%. Isolates 39 high-priority cases of insider information rent extraction. Proposes transitioning to an "agentic regulatory state" — active AI auditing nodes replacing passive data repositories.
  • Relevance to agentic commerce: The "agentic regulatory state" concept — where AI agents serve as autonomous market regulators — is directly relevant to how agentic commerce will be governed. If the SEC/CFTC adopt AI-powered surveillance, it changes the trust calculus for autonomous trading agents. Connects to Chainlink's Taylor Lindman (now SEC Crypto Task Force) and the broader regulatory stack being built around agent transactions.
  • Link: https://arxiv.org/abs/2602.17895
  • Published: 2026-02-19 | Categories: q-fin.CP, q-fin.GN

• "Can Interest-Bearing Positions Solve the Long-Horizon Problem in Prediction Markets?" — Caleb Maresca

  • Abstract summary: Evaluates whether paying interest on locked prediction market positions can fix the "long-horizon problem" (reduced liquidity for distant events). Uses agent-based simulations with LLM traders in a 2×2 factorial design. Finds the long-horizon problem may be overstated — observed pricing bias is only 0.72 pp, much smaller than theoretical estimates. Interest-bearing positions eliminate ~83% of horizon effects and more than triple participation (17% → 62% of wealth). Interest works primarily by incentivizing participation rather than correcting bias.
  • Relevance to agentic commerce: LLM agents as prediction market participants is a preview of autonomous economic agents. The finding that interest-bearing positions dramatically increase participation has implications for how agent payment systems should handle capital lockup — relevant to DeFi liquidity provision where agents like those in the RAmmStein paper (below) are already operating.
  • Link: https://arxiv.org/abs/2602.21091
  • Published: 2026-02-24 | Categories: econ.GN

• "RAmmStein: Regime Adaptation in Mean-reverting Markets with Stein Thresholds" — Pranay Anchuri

  • Abstract summary: Formulates concentrated liquidity provision in DEXs as an optimal impulse control problem. A Deep RL agent learns to manage positions in Uniswap-style AMMs, incorporating mean-reversion speed as input. Evaluated on 6.8M real Coinbase trades at 1Hz. RAmmStein achieves 0.72% net ROI, reduces rebalancing frequency by 67% vs. greedy strategies while maintaining 88% active time. Demonstrates that "regime-aware laziness" improves capital efficiency.
  • Relevance to agentic commerce: This is an AI agent autonomously managing DeFi liquidity positions — exactly the kind of financial activity that lobster.cash, Coinbase Agentic Wallets, and Circle nanopayments are enabling. The agent's ability to learn when NOT to act ("laziness") is a design insight for any autonomous payment agent that must manage gas costs and slippage.
  • Link: https://arxiv.org/abs/2602.19419
  • Published: 2026-02-23 | Categories: cs.LG, q-fin.TR

• "Contextual Memory Virtualisation: DAG-Based State Management for LLM Agents" — Cosmo Santoni

  • Abstract summary: Proposes treating accumulated LLM understanding as version-controlled state, borrowing from OS virtual memory. Models session history as a DAG with snapshot, branch, and trim primitives that enable context reuse across parallel sessions. A three-pass structurally lossless trimming algorithm preserves all content verbatim while reducing tokens by mean 20% (up to 86%). Includes reference implementation for Claude Code. Evaluated across 76 real-world coding sessions.
  • Relevance to agentic commerce: As agents handle increasingly long-horizon commerce tasks (negotiations, multi-step transactions, recurring purchases), context management becomes critical. This DAG-based state management could be the foundation for how agent "memory" persists across commercial transactions — essentially a version-controlled record of agent decisions and commitments. Directly relevant to Claude Code/OpenClaw agent architecture.
  • Link: https://arxiv.org/abs/2602.22402
  • Published: 2026-02-25 | Categories: cs.SE, cs.AI, cs.HC, cs.OS

• "FaMA: LLM-Empowered Agentic Assistant for Consumer-to-Consumer Marketplace" — Yineng Yan et al. (Meta)

  • Abstract summary: Introduces Facebook Marketplace Assistant (FaMA), an LLM-powered agentic assistant that shifts C2C marketplace interactions from complex GUI to conversational AI. Automates high-friction workflows: sellers get simplified listing updates, renewal, and bulk messaging; buyers get conversational product discovery. Achieves 98% task success rate and up to 2× speedup on interaction time. Argues this "agentic, conversational paradigm" provides a lightweight alternative to traditional app interfaces.
  • Relevance to agentic commerce: Meta/Facebook deploying production agentic commerce at scale. The 98% success rate on complex marketplace tasks demonstrates that LLM agents are ready for real commercial transactions. This is one of the largest companies validating the agentic commerce thesis — the shift from GUI-based to agent-based marketplace interaction.
  • Link: https://arxiv.org/abs/2509.03890
  • Published: 2025-09-04 | Citations: 3

📊 Working Papers & Reports (NBER)

• "Building Pro-Worker Artificial Intelligence" — Daron Acemoglu, David Autor, Simon Johnson (NBER w34854)

  • Abstract summary: Three of the most influential economists in AI/labor research define "pro-worker AI" as technology that expands worker capabilities rather than replacing them. They categorize five types of technological change: labor-augmenting, capital-augmenting, automating, expertise-leveling, and new task-creating — only the last is unambiguously pro-worker. They illustrate across aviation maintenance, education, patent examination, and gig delivery. Identify market failures (misaligned incentives, path dependence, "pro-automation ideology") that underinvest in pro-worker AI. Propose nine policy directions including tax code reform, antitrust enforcement, and IP protections for worker expertise.
  • Relevance to agentic commerce: Acemoglu (MIT, recently won Nobel), Autor (MIT), and Johnson (MIT) are the heavyweight trio shaping policy thinking on AI and labor. Their "pro-automation ideology" critique applies directly to the agentic commerce narrative — are we building agents that augment human commerce or replace human workers? Their policy proposals (tax reform, antitrust) will likely influence how governments regulate autonomous commercial agents. This paper will be cited in every major AI policy document for the next 5 years.
  • Link: https://www.nber.org/papers/w34854

• "Chaining Tasks, Redefining Work: A Theory of AI Automation" — Mert Demirer, John J. Horton, Nicole Immorlica, Brendan Lucier, Peyman Shahidi (NBER w34859)

  • Abstract summary: Models production as sequences of steps executed manually, AI-augmented, or fully automated within contiguous "chains." Firms optimally bundle steps into tasks and jobs, trading specialization gains against coordination costs. Key insight: AI "chaining" (where adjacent automated steps create synergies) means comparative advantage logic can fail — some steps are automated not because AI is better at them, but because they're adjacent to other automated steps. The model implies non-linear productivity gains from AI quality improvements. Provides empirical evidence that AI-executed steps co-occur in chains and that dispersion of AI-exposed steps across a job lowers AI execution.
  • Relevance to agentic commerce: Microsoft Research + NBER formalizing how AI agents "chain" tasks — this is exactly what happens in agentic commerce (agent searches → evaluates → negotiates → pays → confirms). The chaining effect explains why end-to-end agentic solutions (like OpenClaw with integrated payments) have an inherent advantage over piecemeal automation. The non-linear productivity gains finding validates the VC thesis that agentic commerce will create outsized value.
  • Link: https://www.nber.org/papers/w34859

• "Public Finance in the Age of AI: A Primer" — Anton Korinek, Lee Lockwood (NBER w34873)

  • Abstract summary: Examines optimal taxation across two stages of AI transformation. In the first stage (AI displaces labor), consumption taxation becomes the primary revenue instrument. In the second stage (AGI systems both produce value and absorb resources), taxing human consumption becomes inadequate. The authors frame AGI taxation as an "optimal harvesting problem" — the tax rate on AGI depends on how humans discount the future. Evaluates specific proposals: taxes on robots, compute, tokens, sovereign wealth funds, and windfall clauses. Shows that differential commodity taxation gains renewed relevance as labor distortions lose their constraining role.
  • Relevance to agentic commerce: If agents are both producers and consumers in the economy, how do you tax them? This paper tackles the core fiscal question of an agent economy. The "optimal harvesting" framework for AGI taxation could apply to how agent transactions are taxed — relevant to the stablecoin and CBDC discussions (Goldstein et al. below) and to Circle's nanopayment infrastructure. Policy teams at Visa, Mastercard, and the Google AP2 consortium should be reading this.
  • Link: https://www.nber.org/papers/w34873

• "What Drives Money Competition: Comparative Advantage in Payments versus Reserves" — Itay Goldstein, Ming Yang, Yao Zeng (NBER w34865)

  • Abstract summary: Studies competition between different forms of money (stablecoins, CBDCs, bank deposits) through the lens of payment vs. store-of-value functions. Central insight: a money that is "too good" as a store of value circulates less as payment because agents prefer to hoard it (Gresham's law formalized). Models equilibria where monies specialize into distinct roles. Counterintuitively, interest-bearing digital currencies (CBDCs) may NOT threaten bank deposits because higher yields raise the opportunity cost of spending, weakening payment adoption.
  • Relevance to agentic commerce: Directly relevant to the stablecoin-for-agents discussion. Circle's nanopayments, USDC on Solana, and the various agent payment rails all compete as "money for machines." The finding that store-of-value properties can undermine payment adoption has implications for whether agents should hold USDC (stable but earns yield) vs. use instant-settlement tokens. The Gresham's law result predicts that in a multi-token agent economy, agents will spend the "bad" tokens and hoard the "good" ones.
  • Link: https://www.nber.org/papers/w34865

• "Machine Learning Meets Markowitz" — Yijie Wang, Hao Gao, Campbell R. Harvey, Yan Liu, Xinyuan Tao (NBER w34861)

  • Abstract summary: Proposes unifying expected return generation and portfolio optimization into an end-to-end ML framework, rather than the traditional two-stage approach (forecast then optimize). Shows that the standard approach fails because it treats forecast errors equally across securities when the investor only cares about precision for assets most important to their portfolio. Each investor gets their own endogenously determined efficient frontier depending on risk preferences, constraints, and friction exposure.
  • Relevance to agentic commerce: As autonomous investment agents proliferate (see "Autonomous Market Intelligence" above), this end-to-end approach is how they should be designed. The personalized efficient frontier concept maps to how each agent's investment behavior should be customized to its owner's risk profile — relevant to Coinbase Agentic Wallets and DeFi agent strategies.
  • Link: https://www.nber.org/papers/w34861

🏛️ Institutions & Labs to Watch

• MIT Economics (Acemoglu, Autor, Johnson): Three papers in this scan alone touching AI and labor. This group shapes global AI policy more than almost anyone in academia. The "pro-worker AI" framing will dominate policy discussions.

• Google Research (Dütting, Paes Leme, Teng): Working on mechanism design for automated bidding in sponsored shopping — the mathematical infrastructure for agentic commerce in search/shopping.

• UC Berkeley / Dawn Song's lab: IMMACULATE paper on verifiable computation for LLM auditing. Song's group consistently produces foundational security work that becomes industry-standard years later.

• Columbia Business School (Allouah, Besbes, Kanoria): The "What Is Your AI Agent Buying?" paper is the most empirically grounded study of agentic purchasing behavior. This group appears to be building a research program around agentic market design.

• Oxford Machine Learning Research Group (Roberts, Zohren): Multi-agent financial trading systems. One of the few groups doing rigorous backtesting with real market data.

• NBER AI Economics cluster (Korinek, Horton, Immorlica, Lucier): Multiple papers this week on AI automation theory, public finance in AI age, and task chaining. This cluster is producing the theoretical foundations that policymakers will use.

📝 Scan Notes

Source Availability

• arXiv: All four queries returned successfully. Total pool: ~65,000+ papers across queries, but heavy overlap. Most relevant papers from last 48 hours concentrated in cs.GT, econ.GN, cs.AI, and q-fin categories.
• NBER: RSS feed returned successfully. Exceptionally strong week — Acemoglu/Autor/Johnson, Korinek, Horton/Immorlica/Lucier, and Goldstein/Yang/Zeng all publishing simultaneously. This is unusual density for AI/econ topics.
• SSRN: Blocked by Cloudflare (403). SSRN continues to be unreliable for programmatic access. Recommend trying browser-based access in future scans.
• Semantic Scholar: First query returned 41 results successfully. Second query rate-limited (429). Found several valuable conference papers not on arXiv, including the IEEE ICAIC 2026 paper on agentic commerce security.

Key Themes This Week

1. Verification as the bottleneck — "Some Simple Economics of AGI" and IMMACULATE both point to verification/auditing as the scarce resource in an agent economy. This validates the entire trust-layer thesis.
2. Agent marketplaces need mechanism design — Donahue/Raghavan (model marketplace), Dütting et al. (shopping auctions), and Allouah et al. (agent purchasing audit) all converge on: naive marketplace design produces bad outcomes when agents are the buyers.
3. "Agentic commerce" is now an academic category — IEEE ICAIC published a paper explicitly titled "Agentic Commerce" + Columbia Business School is building a research program around it. The term has graduated from industry buzzword to academic field.
4. NBER economists going all-in on AI — Acemoglu, Autor, Korinek, Horton, Immorlica, Harvey all publishing in the same week. The economics profession's center of gravity is shifting rapidly toward AI.

Suggestions for Next Scan

• Try SSRN via browser automation instead of web_fetch☆
• Add Google Scholar alerts for "agentic commerce" and "agent marketplace mechanism design"☆
• Monitor arXiv cs.CE (Computational Engineering) category — the SoK paper was cross-listed there☆
• Track the Columbia Business School group (Allouah, Besbes, Kanoria) for follow-up work on ACES framework☆