What Makes Spending on AI Agents Difficult?

Original | Odaily Planet Daily ( @OdailyChina )

Author | Wenser ( @wenser2010 )

As we enter the fourth quarter of 2026, AI Agents have begun to take over consumption execution both online and offline. The explosive popularity of Muse and Grok Bot indicates that once AI Agents gain operational execution rights, they can perform tasks more effectively.

However, few consider the challenges an AI Agent faces when tasked with payments or transactions on-chain: the absence of a private key wallet, who pays for gas, who authorizes transactions, what the upper limit is, and how to navigate compliance windows?

A number of financial infrastructure, public chain, and payment teams are working to answer the question of "how AI Agents spend money," but before that, three more critical questions need to be addressed: Why now? How will machine payments evolve? How far have we come? Odaily Planet Daily will outline the development of related projects in this article to provide some insights.

Key Question One: Why the Second Half of 2026?

This timing has prerequisites and simultaneously meets the following conditions in the second half of 2026:

First, off-chain Agents have become highly productized. The explosive growth of OpenClaw at the beginning of the year has cleared user education barriers, allowing AI Agents to penetrate real scenarios through the iteration of large models, starting to handle paid API calls and offline task execution. However, the demand for "machines needing to pay" generated in off-chain scenarios has long lacked productized supply for on-chain settlement.

Second, public infrastructure for machine payments has begun to take shape. From the x402 protocol to the MPP machine payment protocol built by Tempo, and the operational landing of the Linux Foundation with 40 members, BNB Chain also released Agent Studio in July this year. The improvement of payment standards and tools has made "Agent payments on-chain" practically feasible.

Finally, there is the growth anxiety behind the development of crypto public chains and projects. As narratives like GameFi, SocialFi, DAO, metaverse, and NFTs hit dead ends due to liquidity shortages, user growth traffic is gradually drying up. AI Agents, which are in a period of rapid growth, are expected to become the next source of gas consumption and transaction activity, potentially becoming a key player in the crypto network ecosystem.

From the perspective of the motivations for various parties, the main drivers and potential risks for the transition to an AI Agent economic ecosystem are as follows:

The real needs behind existing Agent projects and applications face a round of screening:

• Real needs validated by the crypto market: settlement (machines paying for services), authorization (security boundaries and limits), identity (verifiable and traceable), reusable Skill functions (validated off-chain).
• Bubble narratives overly amplified by market sentiment: "completely autonomous AI Agent economies," "Agents massively replacing human accounts." In the short to medium term, the account subjects for payments and transactions will still be humans, with AI Agents only playing the role of executors.

Based on the above information, the wave of "AI Agent economic transition" in the second half of 2026 is not merely a pursuit of hot narratives, but a proactive choice and passive demand from crypto public chains, protocols, and projects to extend their lifecycle and gain competitive advantages amidst real needs and bubble narratives.

Key Question Two: How Will Machine Payments Overcome the "First Hurdle of Agent Spending"?

The industry transition has begun, but the challenges faced by the sector cannot be resolved overnight. Although AI Agents have been widely adopted for nearly a year, the so-called "first hurdle of AI Agent spending" presents the following issues from the perspectives of subject, cost, process, and compliance:

Four Hurdles for AI Agents' Autonomous On-Chain Operations

For AI Agents' autonomous on-chain operations, this is not merely an independent act of "executing a purchase," but involves a series of systemic economic behaviors that include identity access, compliance custody, gas execution friction, and execution result evaluation.

In terms of identity access, AI Agents cannot fulfill traditional KYC, bank account setup, filling out transfer forms, anti-money laundering, and counterparty review compliance requirements—without identity, AI Agents cannot enter the financial system;

Regarding compliance custody, who holds the Agent's wallet and private key, how to ensure they are not leaked, and how to avoid issues like prompt injection, erroneous decisions, and overspending due to unlimited retries—without a "safe deposit box," AI Agents are constantly exposed to security risks;

On the matter of gas execution friction, on one hand, gas costs are often priced in native assets of various chains, involving multi-chain execution, which can easily interrupt autonomous processes; on the other hand, the service fees that AI Agents must pay have also caused mismatches in pricing units, making it impossible for machines to complete transactions autonomously—without a universal ecosystem, AI Agent transactions cannot operate smoothly;

In terms of execution result evaluation, based on the characteristics of high-frequency small-value machine economies, the final irreversibility of on-chain result fulfillment, and the asynchronous nature of machine payment execution and fund settlement, related responsibilities, compliance reviews, and the "historical reputation mechanism" and result evaluations all require clear standards—without performance evaluation, AI Agent execution can only remain at the review stage and cannot form an ecological network.

Overview of the Main Discrepancies Among the Three Existing Market Solutions

Currently, the mainstream solutions for machine payments include the following three:

Different solutions correspond to different protocol standards; behind the dispute over machine payment protocol standards lies a more important question—do machine payment actions occur at the "application layer" or the "on-chain protocol layer"? Whoever controls the pricing layer essentially controls the distribution entry of machine payments. This question also temporarily determines whether crypto public chains are "pipelines" or "cash registers."

Therefore, although the x402 protocol has garnered significant market attention, it is still difficult to conclude that the competition has ended. Specifically, the unresolved issues include:

• Has the "two-layer payment structure of gas costs + service fees" been obscured by machine economy narratives?
• Will the payment standard dispute reach a universal solution, or will we have to rely on different on-chain ecosystems to start anew, leading to fragmentation of related machine payment standards?
• Who bears the compliance (KYB / anti-money laundering AML) costs for stablecoin settlements? Is it the on-chain ecological network, the specific transaction parties, or the protocol builders?

According to the latest data released by Chainalysis on October 15, as of the end of the third quarter of 2026, the x402 payment protocol under Coinbase has seen approximately 138.5 million transactions on the Base blockchain; last year in Q3, this number was only 58,800, showing rapid growth.

Key Question Three: How Far Has Agent On-Chain Development Progressed?

Through the development of x402 transactions, we can also gain insights into the four aspects of AI Agent on-chain operations that have achieved certain breakthroughs. Based on their development stages (formal operation stage/demo sample stage/paper planning stage), we can briefly summarize the following information:

Currently, the maturity ranking of industry projects is payment > identity > hosting > authorization. Because it involves red line issues such as security boundaries and user control, the "authorization" part is the first to be disproven in the narrative of "Agent autonomy" and has not yet fully broken through, still relying on demo sample testing from projects like Visa and Mastercard. In terms of hosting and payment, aside from Base, Solana, and Tempo networks, the recently announced Agent Native upgrade by Pharos Network has become an ecological network proposing new solutions.

Pharos Network's Unique Solution: Building a New Ecosystem with Agent Native Network

According to official information, the core positioning of Pharos Network is to achieve the on-chain representation of real-world value through RealFi, with one of the paths being the Agent Native ecological architecture: allowing AI Agents to operate as native economic entities on-chain—holding assets, calling tools, collaborating with other Agents, and paying for their operations.

To achieve this goal, Pharos Network has chosen to build an Agent Native protocol layer, endowing AI Agents with asset, authorization, payment, and identity capabilities, rather than simply adding an AI function to the on-chain application layer.

From the perspective of roadmap planning, Pharos Network's plans are quite steady. Aside from plans to further enhance network infrastructure capabilities in Q4 2026, Agent accounts, transaction interfaces for Agents, State Economy, etc., are still in the protocol definition or testing phase; end-to-end workflows are tentatively scheduled for realization in 2028.

In April of this year, Pharos Network's mainnet Pacific Ocean officially launched; its SALI engine drives deep parallel execution, modular SPN, compatibility with EVM and WASM, and native compatibility with x402 standards, providing convenience for the construction and deep development of the AI Agent economic system; its team members are all former senior executives and core engineers from Ant Group; investors include well-known institutions both inside and outside the industry such as Sumitomo Corporation, Flow Traders, SNZ, Hack VC, and Faction VC.

In addition, its carefully crafted RealFi alliance members include numerous crypto infrastructure service providers such as Circle, Chainlink, Anchorage Digital, Centrifuge, and Morpho. To further incentivize developers and the construction enthusiasm of ecological projects, Pharos Network has also launched a $10 million builder incubation program. Furthermore, in terms of adapting to the AI Agent economic system, the network has made a series of optimizations and preparations:

• Sponsored Transactions (testnet launched): Separating transaction initiators from Gas payers. Applications or sponsors pay Gas, retaining the ownership and control of the initiator, reducing the risk of witch attacks and abuse, while supporting large-scale activities of AI Agents.
• Gas Abstraction / Substitution: AI Agents do not need to purchase Gas midway, supporting uninterrupted execution in multiple steps.
• Agent Accounts & KYA (2027): Account abstraction + permission control framework, maintaining human sovereignty over assets, achieving controllable machine execution.
• AEA (Agentic Economic Activity) framework: Not merely counting Agents or transactions, but measuring effectiveness (meaningful operations triggered by Agents), end-to-end delivery, signal/noise ratio, pipeline continuity, security and self-healing, witch defense, etc.
• High-performance support for micropayments: Official data shows about 30,000 TPS, sub-second finality, compatible with x402 class high-frequency micropayments and protocol operations; supports ERC-20, EVM compatibility, equipped with compliance-related modules.
• Developer and user tools: PharosMCP (on-chain data and operations), SkillEngine, official llms.txt, Port Agent Widget (natural language RealFi interaction, users retain final execution control), ai.pharos.xyz official website.

As AI Agents gradually transform into on-chain economic entities, a single protocol, project, or application can no longer build a complete Agent Ecosystem. Pharos Network constructs the entire circular ecological system while being compatible with existing machine payment networks and protocols.

Open Discussion: The Success and Failure Simulation of the Agent Economic System

After discussing specific solutions, we boldly project which aspects of the future AI Agent economic system are likely to continue developing based on existing information, and which steps face potential risks or the possibility of elimination.

Currently, the aspects with a relatively high probability of success, real demand, and actual trading behavior include: machine payment standards and settlement layers (real payment scenarios exist: API, data, computing power); reusable Skill/tool layers (off-chain verification); KYB-compliant Agent identities (compliance necessity);

The steps most likely to be falsified, leaning towards narrative hype and marketing, are concentrated in three areas: 1) Completely autonomous, unsupervised Agent economies, considering that user control rights cannot be relinquished, this goal is destined to be unachievable in the short term. 2) "Agents massively replacing human accounts," in the medium to long term, the decision-making entity of economic accounts remains human users; as an intermediate step for AI Agent operation, human accounts still hold necessary value and are unlikely to be replaced in the short term. 3) Public chain "AI Agents," this role currently appears to have more marketing attributes than practical value, with no real applications supporting it, and it is difficult to bring real incremental users and usage to the public chain network, making it non-essential at present.

Based on existing information, we can provide a rough timeline inference for the development progress of the AI Agent economic system:

2026-2027, gradual landing period for AI Agent payment and identity aspects;

2027-2028, gradual establishment of AI Agent authorization and collaboration mechanisms;

2028, with the steady advancement of operational authorization processes and compliance supervision, the end-to-end workflow of AI Agents is expected to be further improved and gradually become a key component of the operation of the social economic system. From this perspective, the development roadmap of the Pharos Network Agent Native network considers 2028 as a significant long-term planning milestone.

Of course, the development of the AI Agent economic system is not an overnight process and faces a series of potential risks: such as stagnation in the adoption of the x402 protocol, increasing fragmentation of machine payment standards, liquidity fragmentation across different on-chain ecological networks, significant declines in AI Agent-related transaction volumes, regulatory agencies halting unsupervised on-chain transactions of AI Agents, and leading blockchain networks or ecosystems shifting to new narrative embraces, etc.

Conclusion: The Battle for Entry Points

Returning to the initial question: How do Agents spend money on-chain? The answer likely does not lie in a single technical solution, but in the ownership of entry points------ public chains, HTTP layer payment standards, or on-chain wallets.

Currently, the "ecological-level preparation" provided by Pharos Network may be the best answer, as it represents the "ultimate entry" solution. In the transitional phase of AI Agents from off-chain operational execution to on-chain economic participation, a complete and matching ecosystem encompassing identity, authorization, payment, and carrying is a better answer. The purpose and endpoint of all this align with the trajectory of the influence of AI Agents radiating from the digital world to the physical world------ as Pharos Network's value proposition states, everything is aimed at promoting the operation of real-world value.

Reference materials

x402 protocol transaction volume exceeds 138.5 million: AI Agents reshape crypto payments