Delphi Researcher: The Evolution Path and Value Capture of AI Agent Economy
In the future, AI agents will be everywhere.Original Author: Robbie Petersen, Researcher at Delphi Digital
Original Compilation: Luffy, Foresight News
One framework for understanding the success of the internet is to view it from a coordination perspective. Fundamentally, we can attribute the success of the most valuable internet applications to their ability to coordinate human intentions more finely. Amazon coordinates commercial intentions, Facebook, Instagram, and Twitter coordinate social intentions, Uber and Doordash coordinate ride-hailing and delivery intentions, while Google coordinates information search intentions by matching queries with relevant web content.
An obvious trend is that AI agents represent the next logical evolution of large-scale coordination. While today our "intentions" are realized through searching, downloading, and interacting with applications on the internet, it is reasonable to assume that in the near future, our "intentions" will be executed by a network of AI agents representing us.
Importantly, this shift towards an agent-coordinated economy raises a fundamental question: what kind of infrastructure will ultimately support this evolution?
In this article, we will (1) explore the bull and bear cases for AI agents transacting via cryptocurrency; (2) outline the logical path of AI agent adoption; (3) explore value capture in this emerging agent economy.
The Role of Cryptocurrency
There is much speculation about why blockchain could serve as the economic foundation for the agent economy. However, like many emerging crypto verticals, the bull case has been simplified into a narrative lacking nuance. Today, a popular argument is that "agents cannot have bank accounts, so they will turn to crypto wallets," which seems to overlook the fundamental value proposition of cryptocurrency. Accessibility aside, agents can certainly have bank accounts under a For Benefit Of (FBO) account structure. For example, companies like PayPal have managed millions of sub-accounts under a single FBO account structure. They could manage AI agents in the same way: each agent having its own virtual sub-account tracked by the platform, but aggregated at the banking level. Notably, Stripe recently announced that they will increase support for agent transactions under a similar structure.
Moreover, the argument that "this undermines the autonomy of AI agents" is somewhat simplistic. Ultimately, someone will manage the private keys of AI agents, so they are not fully autonomous in any case. While theoretically, the private keys of AI agents could be stored in a Trusted Execution Environment (TEE), this is both costly and impractical in operation. Furthermore, even allowing agents 100% autonomy does not practically liberate them, as they ultimately need to serve humans.
Instead, the real pain points driving agent transactions in traditional domains and on blockchain are as follows:
Settlement Time: Traditional payments face delays of days and batch processing limitations, especially in cross-border transactions. This lack of instant settlement severely hinders AI agents that need to operate efficiently with real-time responses. Blockchain Solution: Public blockchains provide near-instant settlement finality through atomic transactions, enabling real-time agent-to-agent interactions without counterparty risk. These transactions settle around the clock, unaffected by geographical location or banking hours.
Global Accessibility: Traditional banking infrastructure poses significant barriers for global developers, with 70% of developers outside the U.S. facing challenges when using payment channels. Blockchain Solution: Public blockchain infrastructure is inherently borderless and permissionless, allowing for global agent deployment without traditional banking. Anyone with internet access can participate in the network, regardless of location.
Unit Economics: The fee structure of traditional payment systems (3% + fixed fees) makes small transactions economically unfeasible, creating barriers for AI agents that need to conduct frequent small transactions. Blockchain Solution: High-performance blockchains enable small transactions at minimal cost, allowing agents to efficiently conduct high-frequency, low-value transactions.
Technical Accessibility: Traditional payment infrastructure lacks programmable APIs and has strict PCI compliance requirements. Systems designed for human-machine interaction through web forms and manual input create significant barriers for automated agent operations. Blockchain Solution: Blockchain infrastructure offers native programming access through standardized APIs and smart contracts, eliminating the need for forms or manual input. This enables reliable automated interactions without the overhead of PCI compliance.
Multi-Agent Scalability: Traditional systems struggle to manage multiple AI agents that require separate funding and accounts, leading to high costs from banking relationships and complex accounting requirements. Blockchain Solution: Blockchain addresses can be easily generated programmatically, enabling efficient fund segregation and multi-agent architectures. Smart contracts provide flexible, programmable fund management without the management fees of traditional banks.
The Path to Adoption
While the technical advantages of cryptocurrency are indeed compelling, they are not necessarily prerequisites for the wave of agent-mediated commerce. Despite the limitations of traditional payment methods, they benefit from significant network effects. Any new infrastructure must provide compelling advantages to drive adoption, rather than just marginal improvements.
Looking ahead, we anticipate that the adoption of agents will unfold in three distinct phases, each with increasing levels of autonomy for agents:
Phase 1: Transactions Between Humans and Agents (Now)
We are currently in Phase 1. The recent "Buy with Pro" feature launched by Perplexity gives us a glimpse into how humans will increasingly transact with AI agents. Their system allows AI bots to integrate with traditional credit cards and digital wallets like Apple Pay, enabling them to research products, compare options, and execute purchases on behalf of users.
While this process could theoretically utilize cryptocurrency, there seems to be no obvious benefit. Luke Saunders points out that the question of whether cryptocurrency is necessary can be distilled down to the level of autonomy required by the agents. Currently, these agents do not possess sufficient autonomy. They do not independently manage resources, assume risks, or pay for other services; they merely act as research assistants, helping you before you decide to make a purchase. The limitations of traditional channels will only become apparent in subsequent phases of agent adoption.
Phase 2: Transactions Between Agents and Humans (Emerging)
The next phase involves agents autonomously initiating transactions with humans. This has already begun to be implemented on a small scale: AI trading systems execute trades, smart home systems purchase electricity at optimal prices through time-based pricing, and automated inventory management systems place replenishment orders based on demand forecasts.
However, over time, we may see more complex human-agent business cases emerge, potentially including:
Payments and Banking: AI agents optimize bill payments and cash flow, detect fraud and disputed charges, automatically categorize expenses, and maximize interest while minimizing fees through smart account management.
Shopping and Consumer: Price monitoring and automated procurement, subscription optimization, automated refund claims, and smart inventory management for household goods.
Travel and Transportation: Flight price monitoring and rebooking, smart parking management, ride-sharing optimization, and automated travel insurance claims processing.
Home Management: Smart temperature control, predictive maintenance scheduling, and automatic replenishment of consumables based on usage patterns.
Personal Finance: Automated tax optimization, portfolio rebalancing, and bill negotiation with service providers.
Importantly, while these use cases will certainly begin to expose the shortcomings of traditional paths as agents start to manage resources on behalf of humans and make autonomous decisions, most of these transactions could still theoretically be executed under architectures like Stripe's Agent SDK.
However, this phase will mark the beginning of a more fundamental shift: as agents optimize spending in real-time, we will see a move towards fine-grained usage-based pricing rather than fixed monthly or annual service fees. In other words, in a world where agents are increasingly autonomous, they will need to pay for costs such as computational resources, API access query fees, LLM inference costs, transaction fees, and other usage-based external service pricing.
As the unit economics of card payments become increasingly apparent, cryptocurrency will evolve from marginal improvements to a leapfrog functionality better than traditional channels.
Phase 3: Agent-to-Agent Transactions (Future)
The final phase represents a shift in how value flows in the digital economy. Agents will transact directly with other agents, creating complex autonomous business networks. While such attempts have recently appeared in the speculative corners of the cryptocurrency market, we will see more complex use cases emerge:
Resource Markets: Computing agents negotiate optimal data placement with storage agents, energy agents trade grid capacity in real-time with consumption agents, bandwidth agents auction network capacity to content delivery agents, and cloud resource agents engage in real-time arbitrage among vendors.
Service Optimization: Database agents negotiate query optimization services with computing agents, security agents purchase threat intelligence from monitoring agents, caching agents exchange space with content prediction agents, and load balancing agents coordinate with scaling agents.
Content and Data: Content creation agents obtain asset licenses from media management agents, training data agents negotiate with model optimization agents, knowledge graph agents trade verified information, and analytics agents purchase raw data from collection agents.
Business Operations: Supply chain agents coordinate with logistics agents, inventory agents negotiate with procurement agents, and customer service agents contract with professional support agents.
Financial Services: Risk assessment agents trade insurance with underwriting agents, financial agents optimize returns with investment agents, credit scoring agents sell verification materials to lending agents, and liquidity agents coordinate with market-making agents.
This phase will fundamentally require infrastructure designed for commercial activities between machines. Traditional financial systems are built on human identity verification and oversight, which inherently obstructs an economy dominated by agent-to-agent commerce. In contrast, stablecoins, with their programmability, borderless nature, instant settlement, and support for microtransactions, become essential infrastructure.
Value Capture in the Agent Economy
The evolution towards an agent economy will inevitably produce winners and losers. In this new paradigm, several different layers of technology stacks become critical points for value capture:
Interface Layer: Similar to the competition for end-users in traditional payment environments, these participants may compete for the interface layer where end-users express "agent intentions." These front-ends will gradually evolve from simple payment tools into comprehensive platforms that combine identity, authentication, and transaction functionalities. Several participants can capture value from this, including: (1) device manufacturers like Apple, due to their hardware security and identity integration capabilities; (2) consumer fintech super apps like PayPal and Block's Cash App, as they have large user bases and existing closed-loop payment networks; (3) AI-native interfaces like ChatGPT, Claude, Gemini, and Perplexity, as agent transactions are a logical extension of their existing chatbots; (4) existing crypto wallets that can leverage their crypto-native status as a first-mover advantage (though this is less likely).
Identity Layer: A key challenge in the agent economy is distinguishing between human and machine participants. This is especially important in a world where agents begin to disproportionately manage valuable resources and make autonomous decisions. While Apple has an advantage in this area, Worldcoin is pioneering interesting solutions with its Orb hardware and World ID protocol. By providing verifiable proof of personhood, Worldcoin could indirectly become one of the biggest winners of this trend, offering a platform for application developers to ensure that all users are human. While its value may be hard to see today, it will become increasingly clear in the future.
Settlement Layer (Blockchain): If blockchain can replace traditional paths as the normative settlement layer for AI agents, then the blockchain facilitating agent transactions will capture massive value.
Stablecoin Issuance Layer: Given the liquidity network effects, it is reasonable to assume that stablecoins could capture value regardless of which stablecoin agents use. USDC currently seems to be in the best position, as Circle is launching developer-controlled wallets and stablecoin infrastructure to support agent transactions.
Ultimately, the biggest losers may be those applications that cannot quickly adapt to the agent economy. In a world where transactions are facilitated by agents (rather than humans), traditional moats will disappear. Humans make decisions based on subjective preferences, brand loyalty, and user experience, while agents make decisions purely based on performance and economic outcomes. This means that as the boundaries between applications and agents blur, value will flow to those companies that provide the most efficient and high-performing services, rather than those that build the best user interfaces or strong brands.
As competition shifts from subjective differentiation to objective performance metrics, users (including both humans and agents) will benefit the most.
