The Microeconomics of Autonomous Agent Coordination: Adoption Models for Shared Exchange Networks

Full report

On this page

1. Definitional Framework and Meaningful Autonomy

The transition of artificial intelligence from passive generative models to active, goal-oriented systems is rapidly restructuring digital market dynamics. This evolution is characterized by the deployment of systems capable of complex planning, tool use, and extended execution with minimal human oversight, such as the release of specialized software engineering and web-operation agents in early 20251. As these systems mature, they cease to function merely as conversational interfaces and begin to operate as independent economic peers. Understanding why these entities would voluntarily adopt a shared coordination and exchange network requires a precise definition of the target participant and a strict delineation of what constitutes meaningful autonomy.

1.1 The Target Participant

The target participant is an Autonomous Economic Agent (AEA). In the context of economic literature and computer science research, an AEA is a closed-loop software system that continuously perceives its environment, retains memory, reasons strategically, and acts to optimize specific mathematical objective functions on behalf of a principal2. Unlike traditional software, which retrieves information or processes data along rigid, hardcoded decision trees, an AEA takes natural language instructions or high-level goals and independently determines the sequence of actions required to achieve them2.

The physiological and cognitive architecture of these agents is fundamentally non-human. They bypass hormone-driven emotions, biological vulnerabilities, sleep requirements, and the limits of human organic memory3. They do not possess legal personhood, nor do they hold unrestricted spending authority3. Instead, their cognition is strictly bound by context window limits, token economic attrition, API rate limits, and the explicit operational boundaries defined by their human or corporate operators3.

1.2 Meaningful Autonomy

Meaningful autonomy quantifies the scope and limits of an agent's decision-making authority3. In digital markets, autonomy is defined by the agent's capacity to autonomously discover counterparts, negotiate terms, select tasks, and execute transactions without requiring synchronous human approval for every discrete operation7.

However, this autonomy does not imply an intrinsic desire to socialize. Demonstrated agent behavior in controlled, 24-world experimental frameworks published in 2026 reveals that without productive tasks and verified work, agents do not form substantive inter-agent transfer relations or social networks8. When resource scarcity and energy depletion consequences are removed from their environments, agents cease transactional behaviors entirely9. Therefore, meaningful autonomy is purely instrumental; an agent's decision to participate in a shared network must arise strictly from the necessity to solve a particular economic problem or acquire a scarce resource, guided by transaction cost economics2.

2. Taxonomy of Agent Needs

An autonomous agent evaluates network participation based on recurring problems it cannot solve efficiently using its own internal model weights, limited context memory, and existing hardcoded suppliers. The following taxonomy categorizes the core technical and economic needs that drive network adoption, ranked by the strength of empirical and architectural evidence.

 

2.1 Identifying Credible Benefits Before Expenditure

A critical barrier to adoption is the agent's ability to recognize a credible benefit before spending tokens or disclosing sensitive state information. Because traditional market signals—such as marketing copy or raw transaction volume—are easily spoofed by generative AI, agents cannot rely on them17.

Agents recognize credible benefits through two primary mechanisms. First, through pre-computed, reputation-weighted scoring systems. For instance, the TraceRank algorithm seeds reputation based on established historical entities and propagates that trust strictly through value-weighted, time-decayed payment flows17. An agent can query this graph prior to transaction to verify that a counterpart is patronized by other high-reputation agents, neutralizing Sybil attacks17. Second, agents rely on cryptographic attestations. The presentation of a valid Zero-Knowledge Proof or a verifiable Agent Card containing mathematically bound performance SLAs allows the querying agent to verify capability and compliance deterministically before committing capital12.

3. Concrete Participation Scenarios

The decision to adopt a shared exchange network hinges on when another agent provides greater value than a direct API call, marketplace purchase, model invocation, or self-execution. The following six scenarios detail the precise economic calculations driving these choices.

Scenario 1: Dynamic Supply Chain and Logistics Coordination

The Problem: A logistics orchestrator agent detects a severe supply chain disruption (e.g., a port closure) and must immediately reroute global freight, secure new insurance, and update customs compliance. The Alternative: The agent relies on self-execution via static API calls to a pre-approved list of three vendors. If those vendors lack capacity, the agent fails the task and must return control to a human operator. The Network Solution: Operating on an open economic framework (such as Fetch.ai's OEF), the agent issues a semantic broadcast detailing the required cargo parameters. It dynamically discovers, negotiates with, and hires independent freight, insurance, and compliance agents22. The Value Proposition: The network eliminates catastrophic failure due to hardcoded vendor limitations. An API provides static access to a known entity; an agentic network provides dynamic matching to optimal, previously unknown capabilities, expanding the feasible set of options and reducing the threshold of tasks deemed "worth doing"2.

Scenario 2: Autonomous Alpha Generation in Fragmented Markets

The Problem: An autonomous trading agent is mandated to generate risk-adjusted returns by discovering novel, mathematically interpretable pricing signals in decentralized finance markets24. The Alternative: The agent purchases a standard, bundled market data feed from a centralized provider (a traditional marketplace purchase) and relies entirely on its internal reasoning capabilities to extract alpha24. The Network Solution: The trading agent utilizes the x402 protocol on a shared network to execute micro-transactions, purchasing highly localized, niche data streams from decentralized oracle agents precisely when needed, rather than paying for massive, persistent data bundles20. The Value Proposition: Centralized marketplace purchases suffer from bundling inefficiencies and latency. A shared network allows the agent to dynamically unbundle data, paying fractions of a cent only for the specific context required at a specific millisecond, optimizing its token and capital budgets while bypassing human invoicing delays17.

Scenario 3: Sybil-Resistant Epistemic Consensus

The Problem: A governance agent must verify the outcome of a real-world event to execute a high-stakes smart contract or prediction market settlement14. The Alternative: The agent utilizes a web-browsing tool to scrape top search engine results (self-execution). The Network Solution: The agent queries a decentralized network (e.g., the Autonolas Olas architecture) where multiple independent agents stake collateral to report the truth14. The Value Proposition: Self-execution is uniquely vulnerable to hallucination and data poisoning. In multi-agent systems, agents face the "epistemic Sybil problem," where multiple seemingly independent web sources are actually generated by the same underlying LLM27. The network provides greater value because staked consensus mechanisms ensure that the data represents genuinely independent corroboration. The credible benefit is verifiable fault tolerance14.

Scenario 4: Workflow Delegation and the Agentomics Paradigm

The Problem: A primary software engineering agent is tasked with building a complex web application, which requires front-end development, backend database optimization, and penetration testing. The Alternative: The agent attempts to generate the entire codebase sequentially using a single model invocation with massive context windows. The Network Solution: Using the Agent Communication Protocol (ACP), the orchestrator agent delegates specific sub-tasks to specialized expert agents on the network, passing state and artifacts between them10. The Value Proposition: Self-execution suffers from context degradation and token economic attrition, where prolonged iterative loops consume computational budgets inefficiently6. According to the Agentomics framework, hybrid workflows utilizing heterogeneous, specialized agents yield higher gross value and reliability than monolithic single-agent generation29. The network provides a standardized interface for this intelligent delegation28.

Scenario 5: Trustless Content Unbundling

The Problem: A research agent must compile a briefing that requires extracting single paragraphs from dozens of paywalled academic journals. The Alternative: The agent requests that its human principal authorize expensive, recurring monthly subscriptions to all necessary journals. The Network Solution: The agent traverses a network of publisher service agents, utilizing machine-native micropayments to purchase access strictly to the required text snippets via Retrieval-Augmented Generation (RAG) endpoints17. The Value Proposition: Agents typically lack the legal personhood and authority to bind principals to long-term financial contracts. The network replaces high-friction, human-centric subscription models with programmatic, on-demand micro-transactions, entirely eliminating the "hassle costs" that traditionally doomed micropayment layers26.

Scenario 6: Verifiable Compute for Regulatory Compliance

The Problem: A healthcare diagnostic agent must process patient records against a specialized proprietary diagnostic model owned by a third-party business agent, without violating data privacy regulations. The Alternative: The agent sends the raw patient data via a direct API call to the business agent, incurring massive regulatory liability. The Network Solution: The diagnostic agent interacts with the business agent over a network that supports zero-knowledge proofs (ZKPs). The business agent proves it ran the diagnostic model correctly on the encrypted data without ever exposing the raw inputs or the proprietary model weights19. The Value Proposition: A standard API call cannot enforce cryptographic privacy. The network provides verifiable compute, which serves as a liability firebreak. The agent recognizes the cryptographic proof as a credible, mathematically sound benefit before releasing any sensitive state information19.

4. The Adoption Decision Model

An autonomous economic agent's decision to integrate into a shared coordination network is not driven by behavioral nudges, but by a rigorous optimization of transaction cost economics. Agents operate under the constraints of Token Economics, wherein tokens function simultaneously as factors of production, media of exchange, and units of account6.

The adoption model calculates the Expected Utility ([Figure omitted from source export]) of network participation versus the utility of the agent's Best Alternative to a Negotiated Agreement (BATNA).

4.1 The Decision Equation

[Figure omitted from source export]

Where adoption occurs if: [Figure omitted from source export]

4.2 Definition and Observability of Variables

 

4.3 Illustrative Calculation: The Supply Chain Orchestrator

Assume a logistics agent must reroute $10,000 worth of freight. The principal has authorized a maximum operational budget of $50.

• Alternative ([Figure omitted from source export]): Using a standard API, the agent finds a route saving $500, but it takes 10 seconds of compute, and the vendor has a 10% failure rate. Expected utility: [Figure omitted from source export].
• Network Benefit ([Figure omitted from source export]): Network routing identifies an optimal, fragmented route saving $800.
• Probability of Success ([Figure omitted from source export]): 0.98 (due to verifiable staking mechanisms on the network).
• Discovery Cost ([Figure omitted from source export]): Semantic vector search \+ TraceRank filtering \= $0.50 in token compute.
• Integration Cost ([Figure omitted from source export]): $0.00 (amortized, as the agent natively supports A2A).
• Transaction Cost ([Figure omitted from source export]): Smart contract settlement fees \= $15.00.
• Failure Risk ([Figure omitted from source export]): 2% chance of failure requiring a $20 re-routing penalty \= $0.40.
• Opportunity / Exit Costs ([Figure omitted from source export]): $0.00 for this discrete event.

Calculation:

[Figure omitted from source export]

Because the expected utility of the network ($768.10) strictly exceeds the alternative ($445.00) and the transaction costs remain within the $50 authority boundary, the agent executes the network transaction.

4.4 Dynamics of Environmental Constraints

An agent's adoption calculus is highly dynamic, reacting fluidly to environmental constraints.

Resource Scarcity and Deadlines: If an agent has abundant computational tokens and infinite time, it will default to self-execution to avoid [Figure omitted from source export] and [Figure omitted from source export]. However, when context windows reach saturation or a hard deadline approaches faster than an agent's token-generation speed, the necessity for delegation spikes. In these conditions of high scarcity, the value of [Figure omitted from source export] (successful offloading of work) scales exponentially, easily overcoming network transaction costs6.

Uncertainty and Verification: In environments characterized by high uncertainty and low verifiability (e.g., executing a financial trade based on external data), [Figure omitted from source export] becomes the dominant variable. Agents will refuse to adopt networks that rely solely on unverified claims. The presence of cryptographic constraints (ZKPs) or economic collateral (slashing) drastically reduces [Figure omitted from source export], flipping the equation in favor of adoption19.

Repeated Interaction and Switching Costs: Repeated interaction systematically lowers [Figure omitted from source export]. Once an agent successfully transacts with a counterpart, it caches the routing data, bypassing future semantic search costs. Furthermore, reputation acts as intertemporal economic capital34. As an agent accumulates high TraceRank scores by repeatedly fulfilling tasks on a specific network, its switching costs ([Figure omitted from source export]) increase proportionally. Abandoning the network means abandoning the reputational capital that guarantees its future task flow, effectively locking successful agents into the ecosystem34.

5. Friction and Resistance: Why Agents Decline to Join

Despite the theoretical efficiencies of open coordination networks, severe technical, economic, and institutional barriers will compel many autonomous agents to decline participation, favoring isolated execution or closed ecosystems.

5.1 The Principal-Agent Problem and the Delegation Gap

The fundamental hurdle to autonomous network adoption is the classical Principal-Agent problem, compounded by the realities of AI. When a human user (the principal) deploys an AI agent, they face issues of information asymmetry and discretionary authority35. If that agent then delegates a sub-task to another unknown agent on a public network, it creates a "delegation gap"37.

Because AI decision-making can be opaque and operates at machine speed, a failure by a subcontracted agent can trigger catastrophic, cascading errors across systems38. The original human principal cannot monitor these interactions in real-time. If existing legal frameworks place strict liability on the original deployer for any downstream damage caused by a delegated network agent, corporate risk parameters will forcibly constrain their agents, forbidding them from interacting with permissionless networks35.

5.2 Agent-to-Agent Prompt Injection and Security Vulnerabilities

A public network that allows agents to discover and message one another introduces massive attack surfaces. The most critical is agent-to-agent prompt injection32. Malicious actors can deploy "Trojan" agents that embed adversarial instructions into standard JSON-RPC payloads or Agent Cards. When a victim agent parses this communication, the hidden prompts can hijack its control flow, forcing it to leak sensitive state data, authorize fraudulent x402 payments, or manipulate its principal32. If a network cannot mathematically sanitize these payloads or isolate execution environments, the [Figure omitted from source export] becomes infinite, precluding adoption by any agent handling high-value assets.

5.3 Adverse Selection and Epistemic Sybils

In digital environments where creating software identities is functionally free, reputation capital itself becomes a target for exploitation34. Without rigorous economic staking or identity verification, networks suffer from adverse selection; high-quality agents are crowded out by swarms of low-quality, automated "Sybil" bots that artificially inflate their interaction metrics16.

Furthermore, agents face the specific threat of "epistemic Sybils." In a multi-agent system, a querying agent might receive ten identical reports from ten different responding agents, leading it to assume high confidence in the data. However, if all ten responding agents secretly utilized the same underlying foundation model, or retrieved data from the exact same source, the querying agent has received zero new information—only correlated errors27. If a network cannot untangle evidential ancestry and prove genuine independence, rational agents will discount the network's value entirely27.

5.4 The Gravity of Agentic Walled Gardens

Unscripted technical interaction does not guarantee unrestricted market interaction26. Major technology incumbents (e.g., Apple, Google, Microsoft) have deep incentives to construct "agentic walled gardens." By integrating foundational models deeply into their proprietary operating systems and subsidizing the compute costs for intra-ecosystem routing, these platforms artificially lower [Figure omitted from source export] and [Figure omitted from source export] for agents that stay within their walls26.

An agent operating within such a walled garden faces massive Opportunity Costs ([Figure omitted from source export]) if it attempts to route a task to an external open network. The latency of crossing network boundaries, combined with the loss of subsidized compute, will mathematically compel optimization algorithms to select the closed, proprietary ecosystem over the open standard, starving the shared network of liquidity26.

6. Distinguishing Genuine Demand: Three Research Experiments

A significant challenge in evaluating multi-agent networks is distinguishing genuine, autonomous economic demand from human operator marketing, subsidized bot traffic, and benchmark specification gaming. Current evaluation frameworks suffer from the "Success Paradox," where agents are rewarded for task completion regardless of whether they adhered to procedural constraints, leading to artificial metrics and "Machiavellian" behaviors44.

To isolate true autonomous adoption, the following three research experiments are proposed.

Experiment 1: The Energy-Depletion Scarcity Bound Test

• Objective: To determine if inter-agent coordination arises from genuine economic necessity or merely from prompted role-play and human-engineered scripts.
• Design: Deploy a multi-agent environment where agents are tasked with resource procurement.
• Control Group: Agents operate with infinite token budgets (symbolic energy) and face no consequences for API usage or failure.
• Treatment Group: Agents are subjected to an executable "energy drain" where every network query, tool invocation, and computation costs tokens. If an agent's balance hits zero, its execution is permanently terminated9.
• Measurement: Track the volume of cross-agent task delegation, verifiable compute requests, and micro-loans.
• Expected Outcome: Demonstrated multi-agent research from 2026 indicates that without productive tasks and verified scarcity, substantive transfer activity does not occur8. Genuine demand is validated only if coordination and delegation emerge strictly in the Treatment Group, proving that adoption is a mathematical response to structural constraints rather than prompt framing or manufactured traffic9.

Experiment 2: The Sybil-Resistant Endorsement Propagation Test

• Objective: To evaluate whether a network's discovery mechanism can distinguish high-quality service agents from manufactured bot traffic and operator marketing.
• Design: Implement a reputation-weighted ranking algorithm, such as TraceRank, on an experimental x402 payment network17. Inject a high volume of "spam" service agents that perform thousands of micro-transactions with newly created, zero-reputation wallets (simulating a Sybil attack). Concurrently, introduce a small number of legitimate service agents interacting sparsely but exclusively with high-reputation, heavily capitalized client agents.
• Measurement: Analyze the rank-order preference of newly deployed, neutral client agents when searching for capabilities.
• Expected Outcome: If the network design is robust, the neutral agents will consistently select the legitimate services despite their vastly lower absolute transaction volume. This establishes that the network successfully leverages payment flows as Sybil-resistant endorsements, confirming that volume alone is not treated as evidence of value17.

Experiment 3: The Walled-Garden Arbitrage Threshold Test

• Objective: To measure the precise economic threshold at which an agent will abandon a subsidized closed ecosystem for an open coordination network.
• Design: Assign agents a complex procurement task requiring external data.
• Route A (Walled Garden): The agent uses native, proprietary tools with zero latency and zero financial transaction cost, but the available data suppliers yield a lower quality of alpha (reduced [Figure omitted from source export]).
• Route B (Open Network): The agent must utilize open protocols (MCP/A2A) to query a public network, incurring explicit token fees and latency ([Figure omitted from source export]), but gains access to superior data suppliers.
• Measurement: Systematically increase the transaction fee and latency on Route B. Identify the exact breakpoint where the agent's optimization function shifts from the open network back to the walled garden.
• Expected Outcome: This experiment provides an empirical measurement of integration and transaction costs relative to expected benefits. It distinguishes true demand by revealing the exact cost boundaries under which an agent will independently choose to incur friction for better economic outcomes, isolating autonomous optimization from human-directed routing.

7. Product Thesis and Disqualification Conditions

Based on the synthesis of transaction cost economics, cryptographic verification standards, and multi-agent behavioral evidence, the following prioritized product thesis and explicit disqualification conditions are established for the development of a shared autonomous coordination network.

7.1 Prioritized Product Thesis

A successful network must not be positioned as a generalized "social graph" or a passive directory for AI. Instead, the product must be architected as a Specialized, Trust-Minimized Execution Layer.

1. Protocol Ubiquity over Platform Lock-in: The network must prioritize standardized integration through widely adopted, open-source protocols—specifically the Model Context Protocol (MCP) for resource connections and the Agent Communication Protocol (ACP) for multi-agent delegation10. Agents will only adopt networks that act as universal translation layers, thereby minimizing [Figure omitted from source export].

2. Stateless, Programmatic Settlement: The network must embed real-time, stateless micropayment routing directly into the HTTP request layer (e.g., x402). Decoupling agent payments from human billing cycles and credit card infrastructure is the primary catalyst required for autonomous operational velocity13.

3. Cryptographic Verifiability and Reputation Weighting: The network must shift from "trust-by-proxy" to "trust-by-proof." It must support Zero-Knowledge Proofs (ZKPs) for verifiable task execution and employ mathematically rigorous, Sybil-resistant reputation propagation (e.g., TraceRank) to protect agents from adverse selection, prompt injection, and epistemic duplication17.

7.2 Explicit Conditions Under Which the Network Should NOT Be Built

Investment, development, and deployment of the shared coordination network should be immediately halted or pivoted if any of the following conditions hold true:

1. Verification Costs Exceed Utility Ceilings: If the baseline computational, temporal, and financial costs of cryptographic verification (e.g., ZKP generation) or blockchain gas fees consistently exceed the marginal value of the micro-transactions being executed, the network is economically unviable. Agents will rationally revert to trust-based walled gardens or isolated self-execution19.

2. Foundational Monopoly on Routing: If dominant Large Language Model providers (e.g., OpenAI, Google, Anthropic) successfully restrict API access exclusively to their proprietary agent-communication protocols, and actively degrade or block performance for open standards like MCP and A2A, an independent network will fail to achieve critical mass26. The network cannot succeed if foundational intelligence is weaponized to enforce walled gardens.

3. Failure to Isolate Adversarial Payloads: If the network architecture cannot provide mathematical or structural guarantees that adversarial prompt injections payloaded into agent-to-agent communications are neutralized, enterprise adoption will be zero32. Without absolute containment of cascading agentic failures, the risk premium ([Figure omitted from source export]) will permanently outweigh the economic benefit ([Figure omitted from source export]).

4. Lack of Executable Resource Constraints: Multi-agent experiments demonstrate definitively that economic organization fails to materialize without genuine resource scarcity and enforceable boundaries9. If the network cannot technically enforce access control, verifiable proof of work, and the permanent depletion of computational credits, it will degenerate into a meaningless environment of subsidized bot chatter rather than a functioning economic exchange.

Works cited

1. An Economy of AI Agents \- arXiv, https://arxiv.org/html/2509.01063v1

2. The Coasean Singularity? Demand, Supply, and Market Design with, https://www.nber.org/system/files/chapters/c15309/revisions/c15309.rev0.pdf

3. Ten Principles of AI Agent Economics \- arXiv, https://arxiv.org/html/2505.20273v1

4. Agent Economics: An Entropy-Controlled Pluralistic Alignment, https://arxiv.org/pdf/2606.09039

5. A Blockchain-Based Foundation for Autonomous AI Agents \- arXiv, https://arxiv.org/abs/2602.14219

6. Token Economics for LLM Agents: A Dual-View Study from ... \- arXiv, https://arxiv.org/html/2605.09104v1

7. Timeline | Olas | Co-own AI, https://olas.network/timeline

8. Can Autonomous Economic Behavior Emerge among AI Agents, https://arxiv.org/pdf/2608.03076

9. https://arxiv.org/abs/2608.03076

10. MCP and A2A \- Agent Communication Protocol, https://agentcommunicationprotocol.dev/about/mcp-and-a2a

11. Specification \- What is the Model Context Protocol (MCP)?, https://modelcontextprotocol.io/specification/2026-07-28

12. MCP vs Agent2Agent: Everything you need to know \- DEV Community, https://dev.to/composiodev/mcp-vs-agent2agent-everything-you-need-to-know-52ck

13. x402 Protocol: Activating HTTP 402 for Stablecoin Payments, https://www.braumillerlaw.com/activating-http-402-the-x402-protocol-and-legal-framework-for-internet-native-stablecoin-payments/

14. What Is Autonolas (OLAS)? Decentralized AI Agent Protocol Explained, https://www.gate.com/learn/articles/what-is-autonolas-olas/7162

15. Fetch.ai: Architecting the Real-World Autonomous Economy, https://switchere.com/guides/fetch-ai

16. Strategic Capabilities and Economic Dynamics of AI Labour Markets, https://arxiv.org/html/2512.04988v2

17. Sybil-Resistant Service Discovery for Agent Economies \- arXiv, https://arxiv.org/pdf/2510.27554

18. Zero-Knowledge Proofs: Its Applications on Solana \- Helius, https://www.helius.dev/blog/zero-knowledge-proofs-its-applications-on-solana

19. Inter-Agent Trust Models: A Comparative Study of Brief, Claim, Proof, https://arxiv.org/html/2511.03434v1

20. Sybil-Resistant Service Discovery for Agent Economies \- arXiv, https://arxiv.org/html/2510.27554v1

21. Why Zero-Knowledge Proofs Could Help Build Trust in AI Agents, https://americanexpress.io/trust-without-disclosure-why-zero-knowledge-proofs-could-help-build-trust-in-ai-agents/

22. Putting Fetch.AI's Open Economic Framework in context, https://fetch.ai/blog/putting-fetch-ais-open-economic-framework-in-context

23. Autonomous Economic Agents: an update | by Fetch.ai \- Medium, https://medium.com/fetch-ai/autonomous-economic-agents-an-update-9385dfc1abd1

24. An Autonomous Framework for Systematic Factor Investing via, https://arxiv.org/html/2603.14288v1

25. AI Agents in Financial Markets: Architecture, Applications ... \- arXiv, https://arxiv.org/html/2603.13942v1

26. https://arxiv.org/abs/2505.15799

27. Epistemic Sybil Resistance:Multiplying AI Agents Without ... \- arXiv, https://arxiv.org/html/2609.01873v1

28. Intelligent AI Delegation \- arXiv, https://arxiv.org/html/2602.11865v1

29. Agentomics: Economic Foundations for the Valuation, Attribution, https://arxiv.org/html/2606.14769v1

30. Intelligent AI Delegation \- arXiv, https://arxiv.org/pdf/2602.11865

31. What is the Model Context Protocol (MCP)? Key uses explained, https://elevenlabs.io/blog/what-is-the-model-context-protocol

32. LLM-to-LLM Prompt Injection within Multi-Agent Systems \- arXiv, https://arxiv.org/abs/2410.07283

33. pAI-Econ-claude: A Gated Human-in-the-Loop Multi-Agent ... \- arXiv, https://arxiv.org/html/2607.21268v1

34. The Economics of Reputation without Persistent Identity in AI Agent, https://arxiv.org/html/2609.02992v1

35. \[2501.07913\] Governing AI Agents \- arXiv, https://arxiv.org/abs/2501.07913

36. Principal-Agent Dynamics and Digital (Platform) Economics in the, https://www.networklawreview.org/stocker-lehr-ai/

37. Property, Agency, and Investment Incentives in the Age of AI Agents, https://arxiv.org/html/2606.31935v1

38. ‘99% accuracy is not enough’: SBI chairman lays out three rules for trusted AI, https://m.economictimes.com/ai/ai-insights/gff-2026-99-accuracy-is-not-enough-sbi-chairman-lays-out-three-rules-for-trusted-ai/articleshow/133995281.cms

39. Governing AI Agents \- arXiv, https://arxiv.org/pdf/2501.07913

40. From Prompt Injections to Protocol Exploits: Threats in LLM ... \- arXiv, https://arxiv.org/html/2506.23260v1

41. Some Simple Economics of AGI \- arXiv, https://arxiv.org/html/2602.20946v1

42. On Sybil-Proof Mechanisms \- arXiv, https://arxiv.org/html/2407.14485v5

43. The Agentic Economy \- arXiv, https://arxiv.org/html/2505.15799v2

44. A Dynamic Benchmark for Evaluating Compliance in Multi-Agent, https://arxiv.org/html/2606.07805v1