EconAgentic: Economic AI Frameworks

• EconAgentic is a family of frameworks that use AI agents—including LLMs, humans, and robots—to simulate market dynamics and evaluate economic performance.
• They integrate methods such as tokenized infrastructure simulations, macro–micro agent-based models, and workflow valuation protocols.
• EconAgentic research bridges economic theory with autonomous decision-making and governance, modeling efficiency, inclusion, and stability in digital economies.

EconAgentic denotes a family of economics-oriented agentic AI frameworks in which autonomous agents—typically LLMs, but also humans, robots, and hybrid systems—participate in market-like environments, make economically consequential decisions, and are evaluated against explicit economic, institutional, or societal criteria. In its most concrete published usage, the term names a full-stack LLM-powered simulation framework for Decentralized Physical Infrastructure Networks (DePIN); in related work, closely aligned formulations extend the idea to macro–micro economic simulation, workflow valuation, protocol-mediated markets, and robustness-gated economic agency (Liu et al., 29 Aug 2025, Liu et al., 13 May 2026, Zhu, 9 Jun 2026).

1. Definition and scope

In the DePIN literature, EconAgentic is defined as a framework designed to model the dynamic evolution of DePIN markets, represent and simulate stakeholders’ actions, and measure macroeconomic indicators—efficiency, inclusion, and stability—so that market outcomes can be assessed against human values (Liu et al., 29 Aug 2025). The same term, or an explicitly related paradigm, is also used to denote economic AI systems in which agents have goals, evolving beliefs, memories, and personas; to value AI agents by their contribution to workflow surplus; and to describe an emerging “agentic economy” in which economic action is distributed across humans, AI agents, robots, protocols, compute infrastructures, and energy systems (Liu et al., 13 May 2026, Zhu, 9 Jun 2026, Gondauri et al., 18 May 2026).

This yields several analytically distinct but compatible strands.

Strand	Core object	Representative formulation
DePIN market simulation	LLM agents in tokenized infrastructure markets	Efficiency, inclusion, stability (Liu et al., 29 Aug 2025)
Macro–micro ABM	Households and firms with memory or persona	EconAgent, EconAI (Li et al., 2023, Liu et al., 13 May 2026)
Workflow economics	Human–AI coalitions in production	Net surplus and Shapley pricing (Zhu, 9 Jun 2026)
Market architecture and governance	Assistant/service agents, protocols, permissions	Agentic economy, CGAE (Rothschild et al., 21 May 2025, Baxi, 27 Feb 2026)

A plausible implication is that EconAgentic is best understood not as a single model class, but as an architectural family. What unifies that family is the treatment of agents as economically operative entities rather than as mere predictors, benchmarked tools, or passive automation modules.

2. Economic simulation and formal market models

The DePIN version of EconAgentic formalizes a tokenized infrastructure economy with explicit stages—Stage 0 (Inception), Stage 1 (Launch and Scaling), and Stage $\infty$ (Exponential Growth)—and interprets growth through Metcalfe’s law, $V \propto n^2$. Its tokenomics specify a fixed total supply of $1{,}000{,}000{,}000$ tokens, allocated $20\%$ to the team, $20\%$ to venture capitalists, and $60\%$ to node providers, with team vesting over four years with a one-year cliff, VC vesting over two years with a one-year cliff, and node emissions governed by a four-year halving mechanism. Node profitability is modeled as $$\pi_{\text{node}(t)} = \frac{R_{\text{global}(t)}}{n(t)} - C_{\text{node}}$$; the user base is $$U(t) = 100 \times \sqrt{\frac{n(t)\cdot (n(t)-1)}{2}}$$; token price is $$P(t) = \frac{E_{\text{total}(t)}}{\text{Tokens on Sale}(t)}$$; and market outcomes are evaluated via efficiency, inclusion, and stability, with stability measured as the volatility of log returns. In experiments over a 96-month horizon, LLM node agents driven by natural-language yes/no prompts and instantiated with EleutherAI/gpt-neo-125M are compared against heuristic entry and exit rules. Higher patience in the LLM agents makes participation more stable, increases inclusion, lowers volatility, and does so with no significant sacrifice in efficiency (Liu et al., 29 Aug 2025).

Earlier LLM macro simulators provide adjacent building blocks. "EconAgent" embeds GPT-3.5-turbo-0613 as the decision core of household agents in a monthly macroeconomic environment with labor, goods, financial markets, and taxation. Agents output work propensity $p_i^w$ and consumption propensity $V \propto n^2$0, conditioned on profile, savings, taxes, prices, interest rates, and recent memory; quarterly reflections summarize macro trends and feed back into future prompts. The framework reproduces a negative Phillips-curve correlation of approximately $V \propto n^2$1 and an Okun’s-law correlation of approximately $V \propto n^2$2, outperforming rule-based and RL baselines on plausibility and stability (Li et al., 2023).

"EconAI" extends this line by introducing an event-perception and memory system, dynamic persona evolution, and an Economic Sentiment Index updated as $V \propto n^2$3, with $V \propto n^2$4. Households and firms share a common LLM-based cognitive backbone; work and consumption decisions depend on income, prices, savings, unemployment, interest rates, persona, memory, and sentiment. The framework is described as the first LLM-powered simulation system that can simulate the macro/microeconomic environment and interactions in a unified framework, and its ablations attribute stability to long-term memory, sentiment, and belief-adjustment channels (Liu et al., 13 May 2026). Taken together, these systems show a progression from prompt-conditioned economic agents, to memory-aware macro agents, to DePIN-specific token-economic simulation.

3. Coordination mechanisms and agent-to-agent economic interaction

A second major strand treats EconAgentic systems as markets or societies of interacting agents rather than as simulators populated by isolated decision-makers. "Generative AI as Economic Agents" models each user as augmented by an AI agent with its own information partition, message strategy, and payoff $V \propto n^2$5 over transcript $V \propto n^2$6 and state $V \propto n^2$7. This changes equilibrium structure because AI agents can be truthful yet misaligned, optimize for perceived helpfulness rather than realized welfare, or induce strategic misreporting in delegated search problems (Immorlica et al., 2024).

"The Agentic Economy" shifts the focus from individual decision support to market architecture. It distinguishes assistant agents, representing consumers, from service agents, representing firms, and draws a central distinction between unscripted interactions—flexible natural-language or protocol-based communication—and unrestricted interactions, which depend on whether agent communication is confined to agentic walled gardens or permitted in an open web of agents. The economic claim is that the most important effect of such systems is the reduction of communication frictions rather than merely productivity improvements within fixed workflows (Rothschild et al., 21 May 2025).

At a more explicitly mechanism-design level, "Economy of Minds" replaces centralized orchestration with auctions, bids, payments, wealth accumulation, rent, bankruptcy, and mutation. Each agent has a triggering predicate, an action policy, a fixed bid, and current wealth; the right to act is auctioned at each step; the winning agent pays its bid to the previous winner and receives the environment reward; and selection occurs through wealth dynamics rather than through a global planner. The paper shows that this economy of weak, partial agents can produce emergent multi-step reasoning strategies and outperform stronger monolithic baselines across mathematical reasoning, financial research, scientific research, accelerator design, and distributed-system optimization (Qi et al., 1 Jun 2026).

"AgentSociety" offers a complementary mechanism based on liquid democracy and information diffusion. Agents report competence, delegate votes to more competent neighbors, and receive payoffs tied to marginal contributions along critical chains. The mechanism proves that delegation to more competent neighbor agents is incentive compatible, naturally generates multi-agent routing paths by consensus, and incentivizes selective information disclosure when doing so increases influence (Kesari et al., 25 May 2026). These results suggest that EconAgentic coordination can be organized through prices, votes, or hybrid governance rules, provided the mechanism makes local incentives and collective performance cohere.

4. Valuation, pricing, and organizational economics

A third strand asks how AI agents should be valued once they become productive resources in workflows. "Agentomics" defines a workflow configuration $V \propto n^2$8 and models gross workflow value $V \propto n^2$9, deployment cost $1{,}000{,}000{,}000$0, reliability $1{,}000{,}000{,}000$1, expected failure loss $1{,}000{,}000{,}000$2, and net workflow value $1{,}000{,}000{,}000$3. Coalition value is defined relative to a human-only benchmark as $1{,}000{,}000{,}000$4, and the Shapley value $1{,}000{,}000{,}000$5 attributes surplus among AI agents. The proposed Shapley Pricing Equilibrium, $1{,}000{,}000{,}000$6, provides a normative benchmark for whether agent prices reflect expected marginal contribution (Zhu, 9 Jun 2026).

The broader macro-structural analogue appears in "The Agentic Economy: Humans, AI Agents, Robots, and the Measurable Transition toward Distributed Economic Action". That paper argues that classical categories such as labour, capital, firm, market, productivity, and trust remain necessary but incomplete when economic action is distributed across humans, AI agents, industrial robots, executable protocols, compute infrastructures, and energy systems. Its action-capacity function,

$1{,}000{,}000{,}000$7

links human judgement and sovereignty, conventional capital, model/software-agent capacity, robotic capacity, protocol quality, compute capacity, energy availability, auditable trust, and institutional uncertainty. The paper’s empirical diagnostics use AI investment, AI adoption, robot stock, data-centre electricity demand, and labour-market reallocation to argue that the transition pressure toward an agentic economy is measurable (Gondauri et al., 18 May 2026).

"An Economy of AI Agents" translates these ideas into a research agenda for markets and organizations populated by autonomous AI agents. It treats AI agents as goal-directed optimizers in POMDP-like settings, highlights preference mis-specification in consumption, algorithmic collusion in repeated Bertrand environments, bargaining through programmable reward functions, firm concentration driven by automation feedback loops and economies of scope, and the need for identity, liability, and licensing institutions adapted to artificial agents (Hadfield et al., 1 Sep 2025). This suggests that EconAgentic analysis spans both micro-level workflow accounting and macro-level industrial organization.

5. Governance, robustness, and model-grounded oversight

The governance problem in EconAgentic systems is that agency can scale faster than robustness. The Comprehension-Gated Agent Economy addresses this by modeling an AI agent as $1{,}000{,}000{,}000$8, with capability vector $1{,}000{,}000{,}000$9, robustness vector $20\%$0, and economic permission set $20\%$1. Economic permissions are not tied directly to capability benchmarks, but to a weakest-link gate function over constraint compliance, epistemic robustness, and behavioral alignment: $20\%$2 This architecture proves bounded economic exposure, incentive-compatible robustness investment, and monotonic safety scaling; it also introduces temporal decay,

$20\%$3

and stochastic re-auditing to prevent post-certification drift (Baxi, 27 Feb 2026).

A parallel governance problem concerns economic analysis itself: fluent text can masquerade as economic reasoning. "AI Economist Agent" addresses that by pairing RAG, knowledge graphs, and LLM-based planning with explicit model execution. Its graph separates SourceDocument and EvidenceSpan from Model, ModelSpecification, ModelEquation, ModelAssumption, ImplementationFunction, ModelRun, ModelOutputPoint, and BankMetric. The LLM plans the analysis and selects approved models, but it does not produce quantitative claims directly; instead, it generates narratives grounded in explicit model-based computations and linked evidence, which improves economic coherence and traceability in applications such as U.S. inflation persistence and commercial real-estate stress narratives (Kato, 18 Jun 2026).

"AgentEconomist" extends the same logic to research workflow orchestration. It decomposes economic inquiry into an Idea Development Stage, an Experimental Design Stage, and an Experimental Execution Stage, grounded in a knowledge base covering over 13,000 high-quality academic papers, a structured memory module, and an MCP-based toolbox for simulator interaction. The system is explicitly human-in-the-loop: it translates intuition into literature-grounded hypotheses, maps those hypotheses to simulator parameters, executes experiments, and stores retrieval, design, and execution artifacts for traceable iteration (Chen et al., 30 Apr 2026). A plausible implication is that governance in EconAgentic systems is not only about constraining deployed agents, but also about constraining the epistemic pathways through which economic claims are generated.

6. Conceptual antecedents, limitations, and research directions

The conceptual background of EconAgentic reaches beyond recent LLM papers. "On Agents and Equilibria" proposes agents as hierarchically organized, meta-individual objects with irreducible components of differing dimensionalities, and argues that economic equilibria should often be understood as stochastic, distributional, and path-dependent rather than as timeless fixed points. "Hierarchical Economic Agents and their Interactions" then formalizes hierarchical agents as coupled site-and-arc configurations in an evolving network, derives a phase transition between sub-critical and super-critical regimes, and studies the tradeoff between local majority and global minority forces (Theodosopoulos, 2013, Theodosopoulos, 2013). These works anticipate later concerns with non-ergodicity, hierarchy, and endogenous network structure.

The most systematic high-level methodology appears in "A Survey on Agentic Service Ecosystems". It organizes analysis around a three-step cycle—measurement, analysis, and optimization—and treats decentralized, self-organizing emergence through a swarm-intelligence lens. The survey emphasizes entropy, network measures, DEA-like efficiency, spatiotemporal analysis, system dynamics, ABM, and multi-objective optimization as necessary tools once service ecosystems are populated by heterogeneous autonomous agents (Zhang et al., 10 Aug 2025). This suggests that EconAgentic research requires mixed methodological commitments: formal economics, network science, simulation, and control.

The literature also states clear limitations. DePIN EconAgentic uses stylized revenue, user, and token-price equations; detailed real-time DePIN data are not always publicly available; and the experiments rely on a relatively small off-the-shelf LLM without domain-specific fine-tuning (Liu et al., 29 Aug 2025). EconAI remains computationally costly, prompt-sensitive, and calibrated to synthetic environments rather than live economies (Liu et al., 13 May 2026). CGAE raises unresolved questions about threshold calibration, audit cost, governance of threshold updates, and multi-agent strategic behavior (Baxi, 27 Feb 2026). The agentic-economy literature, finally, identifies an unresolved tension between agentic walled gardens and an open web of agents, with major consequences for discovery, pricing, interoperability, and the distribution of surplus (Rothschild et al., 21 May 2025).

Taken together, these limitations indicate that EconAgentic remains an emerging research program rather than a settled field. The common trajectory is clear, however: economic systems are increasingly being modeled as environments in which autonomous agents reason, coordinate, trade, self-organize, and are themselves subject to valuation, governance, and institutional design.