{"id":"24eaf78d-a498-490a-81bf-0d771234df71","arxiv_id":"2505.15799","paper_version":2,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"An economy where AI agents negotiate for consumers and businesses could reorganize markets, and the open or closed structure of agent communication will determine who benefits.","lead":"This paper argues that the biggest economic impact of generative AI will come from reducing the cost of communication between consumers and businesses, not just from making individual tasks faster. It lays out two possible futures for this agentic economy: closed walled gardens run by a few platforms, or an open web of agents that anyone can join.","discovery_kind":"unclear","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Even granting full adoption and interoperability, the paper's core mechanism conflates communication architecture with market architecture: reducing communication frictions alone does not disintermediate platforms because discovery, trust, and feedback aggregation remain scarce and can…","rationale":"The reader's weakest_assumption concerns whether agents will be adopted and able to interoperate programmatically. That is a real precondition, but it is not the most load-bearing concern because the paper's own text acknowledges ongoing progress on interoperability protocols. Even if every consumer and business deploys interoperable agents, the economic conclusion still depends on whether communication frictions are the decisive friction in the transaction. The paper itself identifies discovery, validation, and feedback as remaining scarce resources, which suggests that the architecture of a matching or ranking layer, not the communication layer, determines market power. This makes the central claim overstated. The critique is internal to the economics of the paper rather than an external empirical objection. I recommend no change to the UNVERDICTED verdict because the paper is a scenario essay and my concern sharpens why its predictions are not established, but does not convert it into a testable claim that could be accepted or rejected on current evidence. The reader and I partially agree: we both flag that the disintermediation narrative depends on conditions that are asserted rather than demonstrated, but the reader emphasizes technical adoption while I emphasize that even under adoption the economic mechanism is incomplete.","tokens_in":5351,"tokens_out":2999,"duration_ms":29150,"concrete_test":"Analytically test the Section 3.1 mechanism with a minimal two-sided market model: consumers and sellers each have access cost c per agent-to-agent communication, and each transaction requires a verification or trust cost v that can be provided either by a platform or through decentralized attestation. Solve the model as c goes to zero while v is held fixed. If the platform retains positive market share or charges positive fees whenever v > 0, then lowering communication frictions alone does not eliminate intermediary power, and the paper's central claim fails. Conversely, if the platform's profit goes to zero whenever c goes to zero regardless of v, the claim survives. This directly tests whether the paper's identified bottleneck is actually the one that determines market structure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's load-bearing causal claim is that with low communication frictions, interoperable assistant and service agents can replace two-sided platforms. Section 3.1 says 'once communication frictions are low enough, interoperable AI agents could eliminate the necessity for two-sided platforms.' Yet the same section concedes that platforms provide value 'via discovery, validation, remediation, and economies of scale,' and Section 3.3 states that 'what matters more is the algorithm that matches assistants to service agents' and that 'the truly scarce and valuable resource... will be high-quality human feedback.' This concession undercuts the central claim: if matchmaking, trust, and feedback aggregation remain scarce, then the economic bottleneck is not communication architecture but the governance of the discovery and ranking layer. 'Unscripted' communication is necessary but not sufficient; the paper's key distinction between unscripted and unrestricted is a distinction about governance and market structure, not about communication protocols. The argument therefore relies on an implicit, unproven assumption that discovery, trust, and remediation can be unbundled and distributed in a web of agents rather than reconcentrated in a new intermediary layer. If this assumption fails, the abstract's conclusion that 'the architecture of agentic communication will determine the extent to which generative AI democratizes access to economic opportunity' does not follow, even under the paper's own most favorable technical scenario.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a conceptual essay arguing that the most economically transformative impact of generative AI will come from reducing communication frictions between consumers and businesses. It introduces a framework centered on assistant agents and service agents that interact programmatically, and draws a key distinction between unscripted interactions (technically enabled) and unrestricted interactions (governed by market structure). It surveys current siloed and end-to-end agents, and sketches two future scenarios: closed 'agentic walled gardens' and an open 'web of agents.' It then discusses implications for advertising, micro-transactions, and unbundling of digital goods, concluding that the architecture of agentic communication will determine the extent to which generative AI democratizes economic opportunity.","tokens_in":5683,"tokens_out":2983,"duration_ms":27069,"significance":"If the central argument holds, the paper provides a valuable framing for an important emerging question: whether AI agents will reorganize digital markets and weaken the position of two-sided platforms. Its main contribution is a taxonomy and scenario analysis rather than empirical evidence or formal theory. The paper is transparently an essay: it presents no data, no formal model, and no falsifiable predictions. Its strengths are that it identifies a real design tension (unscripted vs. unrestricted communication; walled gardens vs. open web) and connects the argument to ongoing technical standards efforts such as AutoGen, MCP, and Agent2Agent. As a perspective piece, it has clear value, but the strong causal claims in the abstract currently exceed what the argument supports.","major_comments":[{"comment":"The paper's load-bearing claim that 'once communication frictions are low enough, interoperable AI agents could eliminate the necessity for two-sided platforms' is directly undercut by the same section's concession that platforms provide value 'via discovery, validation, remediation, and economies of scale.' Lower communication costs address only one component of intermediation; they do not by themselves solve the discovery problem, the trust problem, or the dispute-resolution problem. If these functions remain scarce and valuable, then the bottleneck is not communication architecture but the governance of the discovery and ranking layer. The authors should either soften the causal claim or provide a concrete mechanism explaining how these platform functions can be unbundled and redistributed across a web of agents without reconcentrating power in a new intermediary layer.","section":"Section 3.1"},{"comment":"The argument that advertising will decline because 'attention is a less constrained resource' is weakened by the same section's statements that 'what matters more is the algorithm that matches assistants to service agents' and that 'the truly scarce and valuable resource... will be high-quality human feedback.' These concessions show that the scarce resource in an agentic economy is not communication bandwidth but the ranking, matching, and feedback-aggregation layer. The paper should reconcile these statements with the abstract's conclusion that 'the architecture of agentic communication will determine' the democratization outcome. As written, the communication architecture is presented as the determining factor even though the paper's own analysis identifies attention and feedback governance as at least equally important.","section":"Section 3.3"},{"comment":"The 'web of agents' scenario is asserted to require 'large-scale coordination among many players' to develop standards and protocols, and 'robust mechanisms for discovery, trust, and security.' The paper does not explain how these coordination problems are solved in the absence of a central intermediary. Since the functions that platforms currently perform would need to be recreated somewhere, the open-web scenario may simply move concentration from transaction platforms to an infrastructure or discovery layer, rather than democratizing access. This is a central weakness because the paper's democratization claim depends on the feasibility of the open web of agents. The authors should either propose a governance mechanism for the discovery layer or explicitly acknowledge that the web-of-agents scenario carries its own centralizing tendencies.","section":"Section 3.2"},{"comment":"The paper makes strong predictive statements, such as 'the architecture of agentic communication will determine the extent to which generative AI democratizes access to economic opportunity,' without offering testable implications or empirical evidence. Since the essay explicitly relies on future technological adoption and interoperability assumptions, the causal claims are not currently falsifiable. The authors should either add a set of specific, observable predictions that could distinguish the walled-garden scenario from the web-of-agents scenario, or reframe the paper explicitly as a speculative scenario analysis rather than a claim about what 'will' happen. This would make the contribution more rigorous and the central claims more honest.","section":"Abstract and Sections 1-3"}],"minor_comments":[{"comment":"The phrase 'eliminate the necessity for two-sided platforms as intermediaries all together' contains a typo: 'all together' should be 'altogether.'","section":"Section 3.1, paragraph 2"},{"comment":"The sentence 'We expect more extreme unbundling, rebundling, and new products in a webs of agents where there can be unrestricted communication between the assistant any any service agent' contains two typos: 'webs' should be 'web,' and 'any any' should be 'and any.'","section":"Section 3.5, final paragraph"},{"comment":"The phrase 'whether walled gardens of web of agents scenario triumphs' should read 'whether the walled-garden or web-of-agents scenario triumphs.'","section":"Section 3.2, closing paragraph"},{"comment":"Reference [2] is a self-citation by two of the authors. It is used only to support the uncontroversial claim that generative AI improves productivity, so it is not load-bearing; nevertheless, the authors should verify that the citation is appropriate and not inflated.","section":"References"},{"comment":"The paper would benefit from a single figure or table contrasting siloed agents, end-to-end agents, agentic walled gardens, and the web of agents along dimensions such as interoperability, governance, and control. This would make the taxonomy easier for readers to apply.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This is a well-written and timely perspective essay, but the gap between the abstract's strong causal claims and the argument's own concessions (about discovery, trust, and feedback scarcity) is substantial. The paper is suitable for a venue that publishes speculative/position pieces, but the authors should be asked to temper the deterministic language and to address the governance of the discovery layer. The self-citation issue is minor and not a cause for concern. I would not reject, because the conceptual framework is genuinely useful and the overclaim is fixable within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a position essay, not a research result. It's clearly written and worth engaging, but the central claim is overstated in the abstract. The news is that the unscripted/unrestricted distinction is a genuinely useful lens, and the 'preference economy' idea is a fresh way to describe where value concentrates when attention is no longer the bottleneck.\n\nThe paper does several things well. It connects agentic AI to two-sided markets, switching costs, and bundling in an accessible way. It sets out concrete scenarios—walled gardens vs. an open web of agents—and it spends real time on the technical adoption requirements (interoperability, standards). The section on micro-transactions and unbundling is plausible, even if speculative. The citation pattern is fine; the only self-citation (Cui et al.) is not load-bearing.\n\nThe soft spot is the load-bearing claim that low communication frictions can eliminate two-sided platforms. The paper itself concedes the opposite in section 3.1: platforms also provide discovery, validation, remediation, and economies of scale. And in section 3.3 it says the genuinely scarce resource will be high-quality human feedback for matching. Those two concessions undercut the abstract's 'architecture of agentic communication will determine' conclusion. The stress-test puts it well: communication architecture is not the same as market architecture. Even with perfect interoperability, the discovery/trust layer is a natural chokepoint, and it can reconcentrate in a new intermediary instead of disappearing.\n\nThat said, for an essay the paper is honest about uncertainty in the body, and the framework is coherent. It doesn't offer data or falsifiable predictions, which is fine for a conceptual piece if it doesn't oversell. My main recommendation is to soften the causal language and to address the discovery/trust problem explicitly, rather than treating it as an afterthought.\n\nThis is a good discussion piece for people thinking about AI and market structure. I'd bring it to a reading group, and I'd send it to peer review—it deserves a serious referee even though it needs revision. I wouldn't cite it as evidence for any empirical claim, but I might cite it as a framing reference.","headline":"A clear, honest essay on how AI agents might reshape markets, but the central claim overreaches its own concessions about discovery and trust.","tokens_in":6163,"tokens_out":3684,"would_cite":false,"duration_ms":32928,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that the most disruptive economic impact of generative AI will come from reducing communication frictions between consumers and businesses through AI assistants and service agents that interact programmatically, not from…","keywords":["generative AI","AI agents","agentic economy","two-sided platforms","communication frictions","switching costs","micro-transactions","preference economy"],"falsifier":"If, after several years, leading consumer assistant products still cannot transact directly with service agents operated by other firms, and agent-to-agent transactions remain a negligible share of commerce completed without a human in the loop, the predicted reorganization of markets will not have occurred; one could check this by tracking whether an open agent-to-agent transaction protocol is implemented across major platforms and whether meaningful transaction volume moves through it.","tokens_in":5109,"feed_emoji":"🤖","tokens_out":6398,"duration_ms":53144,"temperature":0.7,"pith_summary":"This paper argues that the most disruptive economic impact of generative AI will come not from making individual humans more productive, but from letting AI agents communicate directly with one another. It imagines an 'agentic economy' in which each consumer has an assistant agent, each business has a service agent, and the two exchange information and negotiate transactions programmatically. Lowering those communication frictions would reduce the cost of switching providers and could make today's two-sided platforms, such as travel, retail, and music intermediaries, less necessary. The authors propose that the architecture of agent communication, whether open like the web or closed like app stores, will determine who captures the gains. This matters because the choices about protocols, standards, and market structure are being made now, before the balance of power is settled.","feed_headline":"AI agents talking directly to each other could reshape markets","feed_subtitle":"Lower communication frictions, not worker productivity, is the big economic payoff of generative AI, the paper argues.","key_machinery":"The mechanism carrying the argument is the pair of assistant agents and service agents that communicate programmatically in unscripted ways: assistant agents act for consumers, service agents act for businesses, and they negotiate and transact directly with each other rather than through fixed web forms or human browsing. The paper's central analytical tool is the distinction between 'unscripted' interaction, which is technically feasible through natural-language and protocol advances, and 'unrestricted' interaction, which is governed by market power, standards, and regulation. This distinction allows the paper to separate technical capability from economic constraint, and to frame the future as a contest between closed 'agentic walled gardens' and an open 'web of agents.'","core_discovery":"The central claim is that unscripted, programmatic communication between assistant agents and service agents will reorganize markets by reducing communication frictions between consumers and businesses. The paper draws a key distinction between 'unscripted' interactions, which are made possible by advances in natural language and protocol design, and 'unrestricted' interactions, which depend on market structure and governance. On this basis it argues that when agent-to-agent communication is cheap and flexible, switching costs fall, the matching and standardization role of two-sided platforms becomes less essential, and new forms of commerce arise: ranking by high-quality consumer feedback rather than attention, micro-transactions instead of subscriptions, and dynamic unbundling and rebundling of digital content. Whether these outcomes materialize depends on whether agents interoperate openly across the economy or are confined to walled gardens controlled by a few dominant providers. The paper's ultimate claim is that this architecture will determine the extent to which generative AI democratizes access to economic opportunity.","pith_inferences":["If the paper is right, competition policy should focus on interoperability standards and access to agent communication protocols, since control over that layer will become the new bottleneck for market power.","A testable extension would run controlled comparisons of consumer choice and prices under agent-assisted switching versus current platform search, measuring whether the assumed switching-cost reduction actually materializes.","The walled-garden/open-web framing suggests that even a formally open protocol could still concentrate discovery in dominant providers, so 'unrestricted' communication alone would not guarantee democratization; the paper's own discussion of discovery layers leaves this tension implicit.","One can watch for a measurable shift toward variable pricing of digital components, such as per-article or per-feature charges negotiated by agents, as a concrete leading indicator of the micro-transaction prediction."],"forward_implications":["If agent-to-agent communication becomes cheap and flexible, two-sided platforms will face pressure as intermediaries, although they may retain value through discovery, validation, dispute resolution, and regulatory compliance.","The contest between walled gardens and an open web of agents will determine whether market power concentrates in a few firms or spreads across the economy, because open interoperation requires shared standards, trust, and security mechanisms.","Advertising and discovery will shift from an attention economy to a preference economy, in which scarce high-quality human feedback becomes the main input for ranking and monetizing service agents.","Agent-handled payments make micro-transactions economically viable, favoring usage-based pricing and one-off transactions over long-term subscriptions and repeat business.","Digital goods will be unbundled and dynamically rebundled by agents, and with micro-payments in place, a compensation-based retrieval-augmented generation ecosystem can produce personalized content while paying creators."],"supporting_citations":[{"why":"Documents productivity gains from generative AI, the baseline that the paper contrasts with its communication-friction argument.","marker":"[1]"},{"why":"Provides the switching-costs theory that motivates why communication frictions prevent consumers from taking better deals.","marker":"[3]"},{"why":"Supplies a technical framework that enables programmatic multi-agent interaction.","marker":"[6]"},{"why":"Supplies a protocol standard for connecting agents to external tools and services.","marker":"[7]"},{"why":"Supplies a protocol designed to allow autonomous agents to communicate directly across organizations.","marker":"[8]"},{"why":"Frames general-purpose technologies as engines of growth, which the paper uses to characterize agentic communication.","marker":"[9]"},{"why":"Defines the economics of two-sided markets, the intermediary model that the agentic economy could disrupt.","marker":"[10]"},{"why":"Provides evidence that language-model agents can negotiate, supporting the feasibility of unscripted agent transactions.","marker":"[11]"}],"fun_headline_variants":["Agent-to-agent talk could topple platform middlemen","Cheap agent chatter redraws the market map","Unscripted AI negotiation: the new market force","When AI agents do the talking, markets reorganize"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on consumers and businesses adopting assistant and service agents at scale and on those agents being able to interoperate programmatically across firms, a premise the paper asserts rather than demonstrates.","fun_headline_variants_meta":{"raw":{"variants":["Agent-to-agent talk could topple platform middlemen","Cheap agent chatter redraws the market map","Unscripted AI negotiation: the new market force","When AI agents do the talking, markets reorganize"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1523,"prompt_tokens":949,"completion_tokens":574,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":511}},"tokens_in":565,"tokens_out":574,"duration_ms":5333,"temperature":1.0,"reasoning_tokens":511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:10:44.084916+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If, after several years, leading consumer assistant products still cannot transact directly with service agents operated by other firms, and agent-to-agent transactions remain a negligible share of commerce completed without a human in the loop, the predicted reorganization of markets will not have occurred; one could check this by tracking whether an open agent-to-agent transaction protocol is implemented across major platforms and whether meaningful transaction volume moves through it.","supporting_citations":[{"cited_title":"Generative ai at work","cited_arxiv_id":null,"evidence_quote":"Documents productivity gains from generative AI, the baseline that the paper contrasts with its communication-friction argument."},{"cited_title":"Competition when consumers have switching costs: An overview with applications to industrial organization, macroeconomics, and international trade","cited_arxiv_id":null,"evidence_quote":"Provides the switching-costs theory that motivates why communication frictions prevent consumers from taking better deals."},{"cited_title":"Autogen, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies a technical framework that enables programmatic multi-agent interaction."},{"cited_title":"Model context protocol (mcp), 2024","cited_arxiv_id":null,"evidence_quote":"Supplies a protocol standard for connecting agents to external tools and services."},{"cited_title":"Google agent2agent, 2025","cited_arxiv_id":null,"evidence_quote":"Supplies a protocol designed to allow autonomous agents to communicate directly across organizations."},{"cited_title":"Bresnahan and M","cited_arxiv_id":null,"evidence_quote":"Frames general-purpose technologies as engines of growth, which the paper uses to characterize agentic communication."},{"cited_title":"The economics of two-sided markets","cited_arxiv_id":null,"evidence_quote":"Defines the economics of two-sided markets, the intermediary model that the agentic economy could disrupt."}],"review_version":1}