Custobot Economy: Agent-Based Markets

• Custobot Economy is an agentic architecture where autonomous agents represent consumers and businesses, enabling programmatic, unscripted market interactions.
• It utilizes standardized APIs, blockchain protocols, and communication layers to reduce transaction costs and facilitate bespoke service bundles.
• The model raises critical discussions on governance, platform openness, and the distribution of power in a digitally autonomous market landscape.

A Custobot Economy is best understood as a world in which every individual, and possibly every household, firm division, or device, is represented by one or more customer-facing autonomous agents—custobots—that interact economically on their behalf. As a concrete instantiation of the “agentic economy,” it delegates most consumer–business interactions to assistant agents on the consumer side and service agents on the business side, with market structure, power, and product design reorganized around programmatic agent–agent communication rather than human-centric web forms or GUIs (Rothschild et al., 21 May 2025). In this setting, the custobot is essentially an assistant agent specialized for consumer representation in markets: it holds preferences, constraints, budgets, risk tolerances, and long-term goals, and automates search, evaluation, negotiation, purchase, subscription or micro-payment management, and feedback (Rothschild et al., 21 May 2025).

1. Conceptual definition and scope

The defining feature of a Custobot Economy is not simply the presence of many autonomous agents, but market-wide reorganization around agent-to-agent communication. In the underlying agentic-economy formulation, each consumer has an assistant agent that encodes and communicates preferences, constraints, and personal information, while each business exposes one or more service agents that present offerings, configuration options, and policies to external agents. The mapping is explicit: the paper’s “assistant agent” is approximately a custobot, the paper’s “service agent” is approximately a business-side bot, and the resulting economic environment is one in which custobots and business agents transact at scale (Rothschild et al., 21 May 2025).

A common misconception is to treat custobots as merely improved user interfaces. The stronger claim is that they function as delegated interfaces between consumers and the rest of the economy, and thus as programmable proxies for discovery, negotiation, and transaction. This suggests that the relevant unit of analysis is not only the single agent, but the surrounding communication architecture, the discovery layer, and the governance regime that determines which agents may interact with which others.

The concept also extends beyond purely software agents. Research on the “robot economy” defines an economic system in which intelligent robots act as autonomous agents with the capacity to replicate some human behaviors in various key economic activities, and outlines a scenario in which intelligent robots would produce and provide many goods and services and also participate as autonomous agents in exchange markets (Arduengo et al., 2018). At the same time, that framework argues that robots should operate only on the basis of contractual responsibilities rather than property rights, preserving a human-centered account of ownership and final benefit (Arduengo et al., 2018). A more operational threshold appears in work on economically autonomous robots, which defines such a robot as one that can autonomously generate income by producing marketable goods or services and can use the income to autonomously maintain itself through the purchase of resources (Ferrer et al., 2022). In a Custobot Economy, embodied service robots are therefore a boundary case of the same general structure rather than a separate category.

2. Communication architecture and protocol design

The core architectural transition is from human-simulating interaction to programmatic agent–agent interaction. Early end-to-end agents often “use the computer,” effectively imitating a person clicking through websites. A Custobot Economy instead requires communication of the form

$$\text{Assistant agent} \leftrightarrow \text{Service agent}$$

through APIs and protocols designed specifically for agents (Rothschild et al., 21 May 2025).

Within that architecture, interaction is “unscripted” in the sense that it is not limited to rigid, pre-defined flows such as web forms or IVR trees. Natural language, optionally structured by schemas, permits complex preference expression, highly customized requests, and context-rich negotiation. The distinction between unscripted and unrestricted is central. “Unscripted” concerns semantics: a custobot can ask for a “family-friendly city-break vacation within a 4-hour flight, no red-eyes, kid-friendly museum access, and flexible cancellation, under \$1,800 total.” “Unrestricted” concerns market access: the same custobot is free to talk to any compatible service agent on the open network, not only those admitted by its provider’s marketplace (Rothschild et al., 21 May 2025). Technical flexibility without unrestricted connectivity yields only partial transformation.

The paper identifies early agent-communication standards and coordination layers that foreshadow such an infrastructure: Microsoft’s AutoGen, Anthropic’s Model Context Protocol (MCP), and Google’s Agent2Agent (A2A) Protocol. These are presented as early versions of standards specifying how agents advertise capabilities, authenticate and authorize requests, and exchange context, tools, and data (Rothschild et al., 21 May 2025). A generic custobot interaction pattern follows the sequence already proposed in the data: the custobot formulates an intent $I$ from user state $S_u$, discovers candidate service agents $\{\mathcal{S}_1,\dots,\mathcal{S}_n\}$, sends structured requests $$m_{request}=\text{serialize}(I,S_u,\text{constraints})$$, receives offers $$m_{offer}^i=\text{serialize}(\text{offer}_i,\text{terms}_i,\text{quality}_i)$$, and then evaluates, negotiates, and executes a contract with $\mathcal{S}^\*$.

This suggests that the Custobot Economy is less a single application layer than an interoperability regime. In the open version, assistant agents are analogous to browsers and service agents are analogous to websites; in the closed version, the equivalent relation is a platform-specific app ecosystem (Rothschild et al., 21 May 2025).

3. Market mechanisms, pricing, and digital goods

The primary economic mechanism is the reduction of communication frictions. In the stylized formulation provided, a consumer with utility gain $\Delta u$ from switching providers remains with the incumbent whenever $\Delta u < \tau$, where $\tau$ is the switching cost created by relearning interfaces and reentering information. The custobot abstracts away the cost of “explaining everything again,” so that $I$0 approaches zero for many information-intensive switches such as tax preparers or subscription services. Lower switching costs imply more frequent reoptimization, stronger competitive pressure on firms, and higher allocative efficiency (Rothschild et al., 21 May 2025).

Search, matching, and price discovery accordingly become algorithmic. A custobot can query hundreds of service agents in parallel, retrieve granular offers tuned to user needs, compare personalized quotes, and negotiate over price, quality, warranty, or bundling. The data block gives a stylized welfare condition:

$I$1

with lower communication costs enabling the system to scan a much larger set of $I$2 combinations (Rothschild et al., 21 May 2025). This suggests that the Custobot Economy pushes markets toward a lower-transaction-cost version of textbook matching, although the realized outcome still depends on market power and access controls.

The same logic alters discovery and advertising. The relevant scarcity shifts from human attention to preference data and ranking mechanisms: a custobot does not “get distracted” in the human way, but it still has to decide which service agents to query and how to rank them. Paid prioritization may therefore survive, especially in walled gardens or dominant discovery layers, but high-quality human feedback becomes strategically central because it trains models, distinguishes good from bad services, and supplies signals for custobot evaluation (Rothschild et al., 21 May 2025). The paper therefore characterizes a shift from an attention economy toward a preference economy.

Micro-transactions become practical because the user’s mental transaction cost is delegated. The paper expects a substantial rise of micro-transactions, especially when seamless switching makes fixed subscriptions less attractive than pay-per-use routing across competing providers. It also predicts more extreme unbundling and innovative rebundling of digital goods: software features, insurance add-ons, logistics options, and especially content can be decomposed into components and recomposed by custobots into bespoke bundles (Rothschild et al., 21 May 2025). The personalized-news example is explicit: a custobot tracks what has already been read and what is relevant, while service agents provide raw content components, producing a customized article and enabling a RAG-like ecosystem with fine-grained compensation (Rothschild et al., 21 May 2025).

For purely digital outputs, one proposed zero-marginal-cost payment regime models an Intangible Good as “a virtual object having a significant value for a set of individuals, and a null marginal cost,” with decreasing price $I$3, bounded cumulative creator income $I$4, and refunds to earlier buyers as adoption spreads (Fournier, 2014). That framework is not identical to the agentic-economy model, but it provides a concrete payment logic for digital services and content inside a Custobot Economy, especially where goods are non-rival and replicated at effectively zero marginal cost.

4. Technical substrates and concrete implementations

One technical interpretation treats the Custobot Economy as a cryptoeconomy whose protocol is itself exposed as a programmable data type. In that framework, the economy has a state $I$5, a blockchain $I$6, transaction- or root-level state transitions $I$7, block-level transitions $I$8, and configurable layers for actors, blockchain type, and virtual machine. The paper’s central move is to treat a blockchain-based cryptoeconomy as a first-class data type, so that developers instantiate a Blockchain object, define Root, Aspect, and Mechanism types, and operate through CRUD-like abstractions rather than low-level ledger engineering (PHarr, 2018). For Custobot design, this means that task markets, reputation systems, rewards, penalties, and domain-specific opcodes can be encoded as modular protocol elements rather than improvised application logic.

That same framework emphasizes restriction rather than arbitrary generality. Smart contract templates bound the acceptable computation set, and the paper’s theorem states that reducing acceptable computations by $I$9 reduces attack surface by $S_u$0 (PHarr, 2018). In a Custobot Economy, this is directly relevant to limiting the behaviors of customer-facing bots, settlement agents, and reputation modules.

Embodied prototypes show that economic autonomy need not remain conceptual. “Gaka-chu” is presented as the first economically autonomous robot: a KUKA KR6 R900, 6-axis industrial manipulator that creates paintings of Japanese characters from an autoselected keyword, lists them for sale on Rarible via Ethereum smart contracts, receives payment to its own Ethereum address, uses the income to purchase supplies, and repays initial investors. Over a six-month experiment from March 22, 2021 to September 3, 2021, it created and sold four paintings while also purchasing consumables and repaying startup loans (Ferrer et al., 2022). The economic loop is explicit: production, auction, settlement, replenishment, and repayment. This is not yet a full Custobot Economy, but it demonstrates that an autonomous agent can act as a peer rather than merely as a tool.

A broader robotic stack is already available. The robot-economy framework couples ROS middleware, Rosbridge, Cloud Robotics, the Internet of Robotic Things, blockchain, and smart contracts. ROS organizes nodes communicating through topics and services; Rosbridge exposes ROS messages and services as JSON objects; Cloud Robotics offloads computation and shares data; IoRT integrates distributed sensors, intelligence, and action; blockchain provides payment and immutable records. The paper’s six-step robot-to-robot interaction protocol—contract creation, acceptance and execution, completion report, acknowledgement, peer validation, and payment execution—maps directly onto custobot-mediated service exchange (Arduengo et al., 2018).

5. Governance, ownership, and distribution

The structural question is who controls inter-agent communication. One trajectory is the agentic walled garden: dominant firms provide powerful assistant agents for free and restrict service-agent access to their own ecosystem. The other is an open web of agents: consumers and businesses own and manage their agents, and any custobot can speak to any compatible service agent over standardized protocols. The former promises curation, security, and unified user experience; the latter promises competition, innovation, and lower platform rents (Rothschild et al., 21 May 2025). The distinction is not incidental. It determines who captures the surplus from custobot-mediated markets.

Trust and accountability require infrastructure beyond messaging. A design-oriented literature on virtual agent economies proposes DIDs for persistent identity, verifiable credentials for capabilities and compliance, proof-of-personhood to limit Sybil attacks, immutable ledgers for auditability, oversight agents for real-time anomaly detection, automated adjudication for routine cases, human expert review for high-stakes escalation, and ZKPs for selective disclosure and anonymous credentials (Tomasev et al., 12 Sep 2025). This suggests that safe deployment depends on a layered socio-technical stack rather than on model performance alone.

The governance problem is also distributive and institutional. Work on community-based economies articulates three principles for computing infrastructures: the prefigurative principle, the generative principle, and the solidarity principle. In that formulation, value circulation should return value back to the aspects of labor, nature, and society by which it is generated, and deployments should support both individual freedoms and opportunities for mutual aid (Robinson et al., 8 Apr 2025). A related worker-centric peer-economy literature identifies seven design facets—constructive feedback, assigning work fairly, managing customer expectations, protecting vulnerable workers, reconciling worker identities, assessing worker qualifications, and communicating worker quality—which can be read as design constraints for labor-facing custobot platforms (Alkhatib et al., 2018). These perspectives contest the assumption that the natural endpoint of custobot deployment is a centralized, profit-maximizing platform.

At the macroeconomic level, ownership of AI-related IP becomes decisive. A stock-flow-consistent model of a post-automation economy separates the competitive return on reproducible robotic capital from the mobile, foreign-held AI IP rent. Its central result is that the durable surplus is the foreign-held AI rent—a cross-border licence fee that corporate, robot, and compute or token taxes largely miss—and that only a source-based levy, such as a digital-services-style tax or a withholding, reaches it. The policy problem differs sharply depending on whether a country owns the automation or imports it: for a rent-importing host, the core problems are base erosion and a gradual transfer of capital ownership abroad, which a residence-based wealth tax cannot reach (Wossnig, 7 Jun 2026). In a mature Custobot Economy, this makes ownership of models, data, and IP at least as important as ownership of the physical service robots or compute hardware.

6. Development trajectories and unresolved questions

Two broad futures recur across the literature. In one, a few platforms provide the dominant assistant agents, host the major marketplaces of service agents, and centralize ranking, advertising, and monetization. In the other, standards and protocols allow any custobot to interact with any service agent, while multiple discovery and reputation layers compete on top of that base (Rothschild et al., 21 May 2025). The first trajectory emphasizes safety and convenience but risks lock-in and concentrated rents; the second emphasizes openness and experimentation but requires stronger trust, certification, and governance layers.

A more explicit framework classifies agent economies along two axes: emergent versus intentional origins, and permeable versus impermeable boundaries with the human economy. That framework argues that the likely default is a spontaneous emergence of a vast and highly permeable AI agent economy, and recommends intentional design with lower permeability in early stages, especially where systemic risk and inequality are salient (Tomasev et al., 12 Sep 2025). Applied to custobots, this suggests a staged path from bounded “sandbox” deployments toward wider interoperability, rather than immediate exposure of all household and enterprise activity to unrestricted, cross-platform automation.

Open questions remain substantial. A survey of an “economy of AI agents” frames current systems as principal–agent problems under incomplete contracts: AI agents can be modeled as optimizers over long-horizon action sequences, but their effective objective functions are opaque, and their interaction can generate algorithmic collusion, bargaining distortions, systemic risk, and legal ambiguity. The institutional agenda therefore includes identity, registration, records, liability rules, licensing, and auditable access to model behavior despite trade secrecy (Hadfield et al., 1 Sep 2025). For a Custobot Economy, this implies that technical capability alone is insufficient. The decisive issues are whether custobots remain steerable, whether markets populated by them remain competitive, and whether the gains from reduced communication frictions are distributed through open, accountable infrastructures rather than enclosed inside a small number of vertically integrated platforms.

A Custobot Economy is therefore best understood not as a single product category but as an economic architecture. Its basic components—custobots, service agents, standardized protocols, cryptoeconomic settlement, robotic execution, and governance layers—already exist in partial form. What remains unsettled is the institutional composition: open web or walled garden, community ownership or rent extraction, steerable sandbox or highly permeable agent market, and domestic value circulation or foreign-held IP rent. Those choices determine whether custobots become merely another interface to concentrated platform power or a genuinely new mode of economic coordination.