← Back to Academic Research

Academic Research Scan — 2026-02-24

2026-02-24

Academic Research Scan — 2026-02-24

🔬 High Priority Papers

NBER Working Papers (This Week's Batch — Top Tier)

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

    • Abstract summary: Defines "pro-worker AI" as technology that makes human skills more valuable by expanding capabilities. Proposes a five-category taxonomy: labor-augmenting, capital-augmenting, automating, expertise-leveling, and new task-creating — arguing only the last is unambiguously pro-worker. Demonstrates through real-world examples (aviation, education, gig delivery) that AI's collaborative potential is underexploited relative to its automation capacity. Identifies market failures — misaligned incentives, path dependence, and a "pro-automation ideology" — causing underinvestment in worker-enhancing AI. Proposes nine policy directions including tax reform, antitrust enforcement, and IP protections for worker expertise.
    • Relevance to agentic commerce: This is the definitive framing paper for whether AI agents replace or augment human economic participation. Directly relevant to how agent marketplaces (lobster.cash, OpenClaw skills) should be designed — do they create new human tasks or automate existing ones? The "new task-creating" category maps perfectly to the emerging agent-human collaboration layer.
    • 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 (Microsoft Research / MIT)

    • Abstract summary: Formalizes how production as a sequence of steps can be executed manually, AI-augmented, or fully automated in contiguous "chains." Firms optimally bundle steps into tasks and jobs, trading specialization gains against coordination costs. The key insight: comparative advantage logic fails with AI chaining — meaning traditional economic predictions about which tasks get automated are wrong. The model predicts non-linear productivity gains from AI quality improvements and shows empirically that (1) AI-executed steps cluster in chains, (2) dispersed AI-exposed steps lower overall AI execution, and (3) adjacency to AI steps increases automation likelihood.
    • Relevance to agentic commerce: This is the theoretical foundation for why agent "task chaining" (like OpenClaw sub-agents, multi-step agentic workflows) produces non-linear value. Directly explains why platforms enabling contiguous agent execution (x402 payment chains, ERC-8004 agent-to-agent handoffs) will outperform piecemeal automation. The failure of comparative advantage logic means market designers can't rely on traditional economic models to predict which commerce tasks agents will capture.
    • Link: https://www.nber.org/papers/w34859

  • Public Finance in the Age of AI: A Primer — Anton Korinek, Lee Lockwood (UVA / U Chicago)

    • Abstract summary: Examines how transformative AI will erode the two main tax bases — labor income and human consumption — that underpin modern tax systems. In Stage 1 (AI displacing labor), consumption taxation becomes the primary revenue instrument. In Stage 2 (autonomous AGI producing most economic value), taxing human consumption becomes inadequate. Proposes framing AGI taxation as an "optimal harvesting problem" where the tax rate depends on human discount rates. Evaluates specific proposals: robot taxes, compute taxes, token taxes, sovereign wealth funds, and windfall clauses.
    • Relevance to agentic commerce: ⭐ THIS IS THE PAPER. Korinek is the leading economist on AI economic transformation. His Stage 2 analysis — where autonomous agents produce and consume resources — is exactly the agentic commerce future. The "token tax" analysis directly applies to x402 micropayments and USDC agent transactions. The sovereign wealth fund proposal connects to how agent-generated wealth should be redistributed. Anyone building agent payment infrastructure should read this.
    • Link: https://www.nber.org/papers/w34873

  • What Drives Money Competition: Comparative Advantage in Payments versus Reserves — Itay Goldstein, Ming Yang, Yao Zeng (Wharton / UCL)

    • Abstract summary: Studies competition between monies serving payment vs. store-of-value functions. Central insight: a money "too good" as store of value may circulate less as payment because agents hoard rather than spend (Gresham's law formalized). Delivers equilibria where monies either specialize into distinct roles or coexist with one dominant. Critically: higher yields on digital currencies can weaken payment adoption by raising opportunity cost of spending. Traditional bank deposits may retain dominance over technologically superior digital alternatives.
    • Relevance to agentic commerce: Directly applicable to the stablecoin-vs-CBDC-vs-bank-deposit competition in agent payments. Explains why yield-bearing stablecoins might actually be worse for agent micropayments (agents would "hoard" rather than transact). Implies that USDC's zero-yield design may be optimal for agent commerce, and that Circle's nanopayments approach (maximizing velocity over yield) is theoretically sound.
    • Link: https://www.nber.org/papers/w34865

arXiv — Directly Relevant

  • Evolution of Fairness in Hybrid Populations with Specialised AI Agents — Zhao Song et al.

    • Abstract summary: Studies fairness in hybrid human-AI societies using a bipartite Ultimatum Game model. Separates humans and AI into proposer/receiver groups (unlike symmetric models). Key finding: "Samaritan AI receivers" (strict gatekeepers) drive population-wide fairness far more effectively than "Samaritan AI proposers" (generous givers). Introduces "Discriminatory AI" that predicts co-players' expectations and offers fair portions only to those with high acceptance thresholds. This Discriminatory AI outperforms both Samaritan types, especially under strong selection, and lowers the critical mass needed for equitable steady states.
    • Relevance to agentic commerce: Maps directly to marketplace design: should AI agents in commerce be generous (offering good deals) or strict (enforcing fair pricing)? The finding that gatekeeping AI drives fairness better than generous AI has implications for agent payment protocols — audit/compliance layers (like Sentinel by Valeo) may be more effective at creating fair markets than generous pricing algorithms. The "Discriminatory AI" concept resembles adaptive pricing agents that modulate offers based on counterparty behavior.
    • Link: https://arxiv.org/abs/2602.18498

  • When Coordination Is Avoidable: A Monotonicity Analysis of Organizational Tasks — Harang Ju

    • Abstract summary: Applies distributed systems theory to organizational coordination, showing coordination is necessary if and only if a task is non-monotonic (new information can invalidate prior conclusions). Maps a classic taxonomy of organizational interdependence onto this criterion. Classifies 65 enterprise workflows (74% monotonic = no coordination needed) and 13,417 O*NET occupational tasks (42% monotonic). Introduces the "Coordination Tax" metric. Multi-agent simulations confirm that 24-57% of coordination spending is unnecessary for correctness.
    • Relevance to agentic commerce: This is the theoretical backbone for autonomous agent architectures. If 42-74% of tasks don't actually require coordination, then fully autonomous agents can handle the majority of commerce workflows without human oversight or agent-to-agent negotiation. Directly informs the design of OpenClaw skill architectures and multi-agent orchestration — the paper provides a decision rule for when agents need to coordinate vs. act independently.
    • Link: https://arxiv.org/abs/2602.18673

  • Mechanism Design via Market Clearing-Prices for Value Maximizers under Budget and RoS Constraints — Xiaodong Liu et al.

    • Abstract summary: Studies mechanism design for auto-bidding markets where buyers have private budgets and Return-on-Spend constraints but public (ML-predicted) valuations. Introduces the "extended Eisenberg-Gale program" with RoS constraints, proves it characterizes competitive equilibrium. The market-clearing mechanism is incentive-compatible for financial constraints and achieves a tight 1/2-approximation of first-best revenue. Provides a decentralized online algorithm with sublinear regret convergence.
    • Relevance to agentic commerce: This is essentially mechanism design for AI agent marketplaces. Auto-bidding agents with budget constraints and RoS targets is exactly how agent commerce will work — agents optimizing purchases within spending limits set by their principals. The decentralized online algorithm is directly applicable to real-time agent-to-agent payment protocols. The IC property means agents have no incentive to misreport their financial constraints — a key requirement for trustworthy agentic commerce.
    • Link: https://arxiv.org/abs/2602.19085

  • Discovering Multiagent Learning Algorithms with Large Language Models — Zun Li et al. (Google DeepMind)

    • Abstract summary: Uses AlphaEvolve (LLM-powered evolutionary coding agent) to automatically discover new multi-agent learning algorithms. Evolves two families: (1) In iterative regret minimization, discovers VAD-CFR with novel volatility-sensitive discounting and warm-start mechanisms that outperform state-of-the-art Discounted Predictive CFR+. (2) In population-based training, discovers SHOR-PSRO with hybrid meta-solver blending Optimistic Regret Matching with smoothed temperature-controlled strategies. Both discovered algorithms use non-intuitive mechanisms that human researchers wouldn't design.
    • Relevance to agentic commerce: Demonstrates that AI agents can design better game-theoretic algorithms than humans. In agent marketplaces, the negotiation and trading algorithms will themselves be AI-generated. This is meta-agentic commerce — agents designing the rules of agent commerce. The AlphaEvolve approach could be applied to discover optimal auction mechanisms for agent-to-agent transactions, potentially superior to hand-designed protocols like x402 or ACP.
    • Link: https://arxiv.org/abs/2602.16928

📄 Notable Papers

  • Janus-Faced Technological Progress and the Arms Race in the Education of Humans and Chatbots — Wolfgang Kuhle (econ.GN)

    • Abstract summary: Models how lognormal wage distributions create exponential returns to skill, driving an "educational arms race" where both humans and AI chatbot/agent developers invest maximally — limited only by borrowing constraints. Shows that technological advances increase GDP and inequality while welfare may decline. In the AI interpretation, firms investing in chatbot/agent capabilities face identical corner-solution dynamics, suggesting an AI capability arms race is economically rational but potentially welfare-reducing.
    • Relevance to agentic commerce: Provides the economic theory behind the AI agent capability arms race we're seeing (OpenAI, Anthropic, Google all racing). Implies that agent marketplace platforms should design mechanisms to prevent welfare-reducing overinvestment in agent capability — e.g., quality floors rather than capability ceilings.
    • Link: https://arxiv.org/abs/2602.19783

  • Scaling Inference-Time Computation via Opponent Simulation: Enabling Online Strategic Adaptation in Repeated Negotiation — Xiangyu Liu et al. (cs.MA)

    • Abstract summary: Embeds smooth Fictitious Play (sFP) game theory into LLM inference for repeated negotiation. Uses an auxiliary opponent model that learns in-context to imitate time-averaged opponent behavior, combined with best-of-N sampling against the model. Achieves significant performance improvement in repeated negotiation games without parameter updates — purely through scaled inference-time computation.
    • Relevance to agentic commerce: This is how AI agents will negotiate in marketplaces — building opponent models during repeated transactions and adapting strategy at inference time. Directly applicable to agent-to-agent price negotiation, contract terms, and multi-round procurement scenarios. The "no parameter updates" aspect means deployed agents can improve their negotiation performance purely through experience.
    • Link: https://arxiv.org/abs/2602.19309

  • Stability Anchors and Risk Amplifiers: Tail Spillovers Across Stablecoin Designs — Wenbin Wu (econ.GN)

    • Abstract summary: Analyzes systemic risk across 8 major stablecoins (2021-2025) using Quantile VAR. Three key findings: (1) Fiat-backed stablecoins are "stability anchors" with near-zero spillovers; algorithmic/crypto-collateralized become "risk amplifiers" in extreme conditions. (2) Theoretical risk isolation between fiat and crypto markets breaks down during stress — direct USD-BTC volatility channels emerge. (3) Algorithmic stablecoins show significant residual contagion while fiat-backed show flight-to-quality. Recommends 2-3x higher regulatory capital buffers for non-fiat-backed stablecoins.
    • Relevance to agentic commerce: Agent payment infrastructure (Circle USDC, lobster.cash) relies on stablecoin stability. This paper confirms that fiat-backed stablecoins (USDC) are the safest rails for agent commerce, validating the ecosystem's current trajectory. The 2-3x capital buffer recommendation could inform agent wallet reserve requirements in the Sapiom/XKOVA "Know Your Agent" frameworks.
    • Link: https://arxiv.org/abs/2602.18820

  • RAmmStein: Regime Adaptation in Concentrated AMMs — Optimal Impulse Control — Pranay Anchuri et al. (cs.LG / q-fin.TR)

    • Abstract summary: Applies deep RL to liquidity provision in decentralized exchanges (Uniswap v3-style concentrated AMMs). Formulates position management as an optimal impulse control problem, using regime detection (mean-reversion speed of OU process) as input. Achieves 0.72% superior net ROI by reducing rebalancing frequency 67% while maintaining 88% active time. Key insight: "regime-aware laziness" preserves returns that would otherwise be eroded by gas fees and slippage.
    • Relevance to agentic commerce: Demonstrates autonomous DeFi agents managing real capital with RL. The "regime-aware laziness" principle applies broadly to agent commerce — agents should minimize transaction overhead by being strategically inactive. Directly relevant to agent-operated DeFi positions and the kind of autonomous financial agents that lobster.cash/Skyfire are building.
    • Link: https://arxiv.org/abs/2602.19419

  • A Potentialization Algorithm for Games with Applications to Multi-Agent Learning — Sharwin Rezagholi (cs.MA / cs.GT)

    • Abstract summary: Introduces an algorithm that converts any normal-form game into an approximating potential game with ordinal potential function. Because potential games have well-understood convergence properties, this equips every game with a surrogate reward structure enabling efficient multi-agent learning. Numerical simulations with replicator dynamics confirm convergence to stable behavior.
    • Relevance to agentic commerce: If you can convert any multi-agent interaction into a potential game, you can guarantee convergent agent behavior in any marketplace. This is a foundational technique for ensuring that agent-to-agent commerce systems reach stable equilibria rather than oscillating or collapsing — a key requirement for production agent payment systems.
    • Link: https://arxiv.org/abs/2602.18925

  • LMFPPO-UBP: Local Mean Field PPO with Unbalanced Punishment for Spatial Public Goods Games — Zhaoqilin Yang et al. (cs.GT)

    • Abstract summary: Tackles cooperation in spatial public goods games using deep RL with a reformulated mean field approach. Integrates unbalanced punishment that penalizes defectors proportionally to local cooperator density. Achieves rapid and stable global cooperation even under low enhancement factors, lowering the cooperation threshold below traditional methods.
    • Relevance to agentic commerce: Public goods problems are central to agent marketplaces — maintaining shared infrastructure, reputation systems, and common resources. The "unbalanced punishment" mechanism could inspire penalty structures in agent commerce protocols where defecting agents (e.g., failing to honor x402 payments) are penalized proportionally to the cooperative density of their neighborhood.
    • Link: https://arxiv.org/abs/2602.18696

  • Integrating Predictive Models into Two-Sided Recommendations: A Matching-Theoretic Approach — Suguru Otani et al. (econ.GN)

    • Abstract summary: Models two-sided platform recommendations as a many-to-many matching problem. Introduces "exposure-constrained deferred acceptance" (ECDA) that limits receiver exposure by expected likes/dates rather than headcount. Using data from a large Japanese dating platform, shows ECDA increases effective matches and receiver-side quality despite reducing total matches. Large-scale field experiment confirms the effects.
    • Relevance to agentic commerce: Two-sided matching is the core problem in agent marketplaces — matching service-providing agents with service-consuming agents. The ECDA algorithm addresses the "congestion problem" where popular agents get overwhelmed. Directly applicable to designing fair agent discovery and recommendation in platforms like ClawHub (7,779+ skills).
    • Link: https://arxiv.org/abs/2602.19689

📊 Working Papers & Reports

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

    • Abstract summary: Argues that the traditional two-stage portfolio approach (forecast returns → optimize) is "deeply problematic" because it treats cross-sectional prediction errors as equally important. Proposes an end-to-end ML framework that unifies return generation and portfolio optimization. Each investor gets their own endogenously determined efficient frontier depending on risk preferences, constraints, and friction exposure. Empirical evidence shows the unified approach outperforms two-stage.
    • Relevance to agentic commerce: As AI agents manage portfolios autonomously, this end-to-end approach becomes the standard. The "personalized efficient frontier" concept maps to how each agent should optimize its own transaction strategy based on its principal's constraints — no universal optimal strategy exists.
    • Link: https://www.nber.org/papers/w34861

  • Could Large Language Models Work as Post-hoc Explainability Tools in Credit Risk Models? — Yiqing Wang et al. (q-fin.RM)

    • Abstract summary: Evaluates GPT-4-turbo, Claude Sonnet 4, and Gemini-2.0-Flash as explainability tools for credit risk models. Strong evidence for the "translator" role (converting model outputs to natural language). Autonomous explanation shows low alignment with model attributions. Few-shot prompting helps for linear models but not XGBoost.
    • Relevance to agentic commerce: As AI agents make financial decisions, explainability becomes a regulatory requirement. The finding that LLMs work as "narrative interfaces" but not autonomous explainers suggests agent commerce platforms will need separate audit layers (like 8004scan, Sentinel) rather than relying on agents to explain their own decisions.
    • Link: https://arxiv.org/abs/2602.18895

  • Fiscal Limits to Protectionism: The 2025 U.S. Tariff Laffer Curve — Jack Rossbach et al. (econ.GN)

    • Abstract summary: Quantifies the tariff Laffer Curve using a multi-sector Ricardian model calibrated to the 2025 US trade war. Revenue-maximizing tariffs: 20-30%. Welfare-maximizing: 0-10%. By January 2026, 20% of US tariffs exceed their Laffer peaks (revenue-decreasing). Coordinated retaliation sharply erodes welfare.
    • Relevance to agentic commerce: Trade war dynamics affect cross-border agent commerce infrastructure. If tariffs increasingly target digital services/compute, agent payment protocols may need jurisdiction-aware routing. The finding that many tariffs are past their revenue peak suggests economic instability that could accelerate migration to decentralized agent-to-agent commerce channels.
    • Link: https://arxiv.org/abs/2602.18938

  • Computational Social Choice: Research & Development — Niclas Boehmer et al. (cs.GT, AAMAS '26 Blue Sky)

    • Abstract summary: Calls for "COMSOC-R&D" — a problem-driven research agenda to design, implement, and test collective decision-making systems in the real world. Argues computational social choice has produced sophisticated theory but insufficient deployment.
    • Relevance to agentic commerce: Agent governance (how groups of agents make collective decisions) is an unsolved problem. This AAMAS '26 paper signals the field is pivoting from theory to deployment — relevant to designing governance mechanisms for agent DAOs, multi-agent collectives, and decentralized agent commerce protocols.
    • Link: https://arxiv.org/abs/2602.20074

  • NBER: Venture Fraud — Alexander Dyck, Freda Fang, Camille Hebert, Ting Xu (w34868)

    • Abstract summary: First comprehensive analysis of VC-backed startup fraud (614 cases since 2000). VC-backed firms are 54% more likely to face fraud charges. Founder-controlled boards are 88% more likely to commit fraud. Fraudulent entrepreneurs continue founding new startups unharmed — suggesting lack of market discipline.
    • Relevance to agentic commerce: As AI agents operate startups and manage capital, the fraud vectors shift. The finding that governance > founder characteristics in predicting fraud supports the need for strong agent governance frameworks (KYA, ERC-8004 reputation, agent auditing). The "lack of market discipline" finding suggests on-chain reputation systems (8004scan) could provide the accountability that traditional VC markets lack.
    • Link: https://www.nber.org/papers/w34868

🏛️ Institutions & Labs to Watch

  • MIT Economics (Acemoglu, Autor, Johnson) — Continuing to lead on AI labor economics. Two major papers this week alone. Their "pro-worker AI" framing is becoming the policy default.
  • Microsoft Research (Horton, Immorlica, Lucier) — The "task chaining" theory is fundamental. Horton's experimental economics work is increasingly AI-focused.
  • Korinek Lab (UVA) — Anton Korinek is becoming the economist for AGI-era public finance. His "optimal harvesting" framing for AGI taxation will be influential.
  • Wharton (Goldstein, Zeng) — Strong on digital money competition theory. Watch for applications to agent payment rails.
  • Google DeepMind (Zun Li et al.) — AlphaEvolve discovering game-theoretic algorithms signals DeepMind is investing heavily in multi-agent economics.

📝 Scan Notes

  • arXiv API: Rate-limited (429) on all 4 queries. Fell back to scraping listing pages + individual abstract pages. Need to space API calls further or get institutional access. Consider a 60-second delay between API calls for next scan.
  • Semantic Scholar API: Also rate-limited (429). Should apply for an API key: https://www.semanticscholar.org/product/api#api-key-form
  • Brave Search API: Not configured — still deferred. Would help find papers via search when APIs are down.
  • SSRN: Cloudflare-protected (403). Browser automation required for access — consider adding to the nightly browser session.
  • NBER RSS: ✅ Fully functional. This week's batch was exceptional — 3 directly on-topic papers from top researchers.
  • Overall quality: OUTSTANDING week. The Acemoglu/Autor + Horton/Immorlica + Korinek trio constitutes a near-complete theoretical framework for AI agent economics. Sir should flag these for the portal.
  • AAMAS 2026 papers appearing: Multiple papers tagged "AAMAS 2026" (2602.18506, 2602.20074). The conference proceedings will be worth a dedicated scan when released.
  • Suggestion: Apply for Semantic Scholar API key and set up a 3-second delay between arXiv API calls to avoid rate limiting.