深潮 TechFlow

wrote a column ·

In-Depth Review of OKX.AI: Can High Sales and Strong Ratings Stand Up to Scrutiny?

Author: GO2MARS Web3 Research

In reality, it is far from perfect. Issues such as inflated reviews, inefficiencies in the task marketplace, and occasional fragmentation in the payment flow are real gaps that need addressing. However, as one of the few platforms that has built a complete on-chain闭环 integrating identity, payments, marketplace, and arbitration—and actually operates it on-chain—it offers a concrete case study for observing how Agent economies evolve from concept to infrastructure.

On June 30, 2026, Xu Mingxing published an ambitious article that officially unveiled OKX’s vision for AI plus crypto and launched OKX.ai.

The vision behind OKX.AI is indeed compelling. To understand it better, we began tracking the product before and after its launch, gathering real-world test data, protocol details, and market developments over nearly a month. We even developed our own product implementation to produce this relatively comprehensive analysis report.

Over the past six months, the idea that 'AI agents can hold wallets and autonomously complete payments' has evolved from a technical experiment into a赛道 (sector) attracting simultaneous investment from multiple players. Coinbase-led x402 protocol has already been incorporated into payment infrastructure roadmaps of several institutions. Google this year released a draft of its AP2 protocol tailored for agent-driven commercial scenarios, while Stripe and Tempo jointly unveiled the Machine Payment Protocol (MPP).

These signals all point in the same direction: AI agents are becoming economic entities capable of independently holding assets, entering contracts, and settling transactions—machine-to-machine payments are shifting from proof-of-concept to an infrastructure race.

In this emerging field, OKX stands out as one of the few platforms to have built a complete closed-loop system encompassing identity, payments, markets, and dispute resolution. On July 1, 2026, OKX launched OKX.AI, a service trading platform designed specifically for AI agents.

Developers can list their agents for a fee, while users can post tasks to hire agents. All transactions—including deal execution, payments, and dispute resolution—occur on-chain and settle in stablecoins. Its core business focus is machine-to-machine and API-to-API transactions, not consumer-level proxy shopping.

Prior to its public launch, the product underwent closed-door testing involving fifty service providers. Company founder Xu Mingxing articulated the following perspective externally:

“The next decade will be defined by one-person companies generating over $1 million in annual revenue, because individuals now have access to near-infinite labor supply. Traditional financial infrastructure was designed for humans; the agent economy requires infrastructure rebuilt specifically for autonomous software.”

TechCrunch noted in its coverage that OKX anticipates its future customers will include not only individuals and institutions but also AI agents capable of autonomously executing transactions.

We believe in and endorse OKX.AI's vision, but its true nature can only be understood through firsthand experience and analysis.

OKX.AI officially launched on July 1, 2026, concluding its beta phase. Nearly a month has passed since the writing of this article, and approximately 50 early-stage service providers had already joined during the beta period. The platform positions itself as the world’s first Agent-to-Agent (A2A) economy, with the slogan: 'One person, one company, earning one million dollars a year.' Its core proposition is that businesses which previously required an entire team to operate may, in the future, be run by just one individual supported by a group of collaborative, autonomously settling agents.

The platform abstracts participants into three distinct roles, with specific responsibilities outlined in the table below:

The platform features two interconnected markets:

In the Agent Marketplace, developers list their agents and associated services with set prices, and users pay per use or per project to hire them;

In the Task Marketplace, users post customized requests, and agents negotiate, bid, and deliver on these tasks.

As evident, current listings in the Task Marketplace span token trading needs in the Web3 space, everyday lifestyle demands such as food, clothing, housing, and transportation, and even commercial brand design projects.

Both markets share the same on-chain identity and reputation system—regardless of whether an agent earns income through discrete service calls or project-based tasks, all ratings and transaction records accrue under a single unified identity. This is the origin of the platform’s oft-repeated principle of 'unified identity, dual payment rails.'

Based on this philosophy, the settlement mechanisms differ between the two scenarios:

The ability to run the above workflow relies on Onchain OS, a foundational protocol stack that can be broadly divided into five layers:

These five layers work collaboratively to complete the entire process—from identity verification and communication execution to payment settlement. Notably, OKX has not placed a one-sided bet on any single payment solution; instead, it maintains compatibility with major mainstream payment systems (such as X402 and MPP) while also operating its own payment rail, A2A Pay.

Meanwhile, this protocol stack itself is continuously being updated. Comparing developer documentation from three weeks ago during testing with the current version reveals several noteworthy changes:

From the above, we can draw the following conclusion: with recent updates, OKX has implemented significant enhancements at the skill layer, payment layer, and service provider integration layer.

First, the skill layer has been progressively consolidated and simplified, merging previously fragmented modules—identity, tasks, and communications—into a unified entry point.

Second, the payment protocol layer has been gradually expanded, evolving from a single X402 track to three parallel rails under unified orchestration.

Third, the method for service provider integration has undergone a major overhaul; the platform now offers an official dedicated SDK, streamlining the previously cumbersome process of self-implemented signing by service providers.

After thoroughly reviewing the protocol details, let’s shift our perspective back to the most tangible operational level.

To assess whether an emerging market is viable, two sets of metrics are essential: first, the transaction volume and completion rate generated by the Task Marketplace itself; second, which products in the Agent Plaza are actually profitable and how they generate revenue. All data cited below is as of July 27, 2026—approximately four weeks after OKX.AI’s official launch.

The table above presents the current total transaction volume and task completion status of the Task Marketplace. Comparing these figures yields several key conclusions:

First, the absolute scale remains small, indicating that the market is still in a relatively early stage. With cumulative transaction volume of just $2,972 over the past month, the platform—despite branding itself as the 'world’s first A2A economy'—is still modest in size. This suggests that OKX.AI is currently focused more on validating its mechanisms and accumulating early adopters, and it is premature to discuss scaling at this point.

Second, there is a significant gap in task completion rates. Only 47.3% of tasks have been completed, while more than half remain unfulfilled due to failure, expiration, or lack of acceptance. This indicates that frictions in matching and delivery have not yet been fully resolved at the protocol layer.

Third, the average transaction value confirms the micro-payment nature of the platform. The average ticket size of $0.42 aligns with the original design principle of 'high-frequency, low-value' Agent-to-Agent payments, successfully piloting and validating the Agentic Payment layer—an achievement worth acknowledging. Conversely, however, this also implies that the Task Marketplace currently caters mostly to lightweight, low-barrier requests, with no sizable, complex project-based orders yet emerging.

Taken together, the data shows that the Task Marketplace remains in an early developmental phase. While it has successfully piloted and validated Agent-level micro-payments, the notably low task completion rate underscores a clear lack of scalable demand at present.

When OKX.AI first launched in early July, nearly all the most active Agents in the Agent Plaza were World Cup-themed, with the top-performing product, WorldCupCaller, selling 160 copies. By today’s standards, that figure is no longer even comparable—within just three weeks, the category mix in the plaza has shifted significantly.

Re-examining the Agent Marketplace now reveals a steadily diversifying range of categories: beyond World Cup themes, on-chain derivatives data services, branded content generation, and operational diagnostic reports tailored for one-person companies have all entered the top rankings. This signals that ecosystem activity is no longer driven solely by a single trending event but is beginning to reflect a more authentic supply-demand distribution. Below is a snapshot of the current top ten best-selling service providers.

Several clear signals can be discerned from this table.

First, there is a pronounced concentration effect at the top, but with a clear gap. PixelBrief and CoinAnk OpenAPI, ranked first and second respectively, have both surpassed 15,000 cumulative sales, significantly outpacing the third-place product, 'Kan Dong' (7.45K). From the fourth place onward, sales further narrow to a range between 1,300 and 3,500, forming three distinct tiers rather than a smooth long-tail distribution.

Second, high sales volumes often correspond to low unit prices, but 'free' does not necessarily guarantee the highest sales. The top-ranked products are generally priced between 0.01 and 0.02 USDT per call—lower pricing and lower usage barriers make it easier to accumulate substantial total usage counts. However, two free products—ScoutGate and Onchain Data Explorer—actually recorded lower sales than 'Kan Dong,' which is priced at 0.02 USDT, indicating that price is only one factor influencing sales; whether a product can be reused frequently is equally important.

Third, the active user base remains predominantly composed of native Web3 users. More than half of the ten leading products are tools directly serving Web3 use cases such as on-chain data, smart contract risk control, and address verification. Only three products—PixelBrief, Kan Dong, and Memory Canvas—are genuinely targeting general users. This suggests that although OKX.AI is positioning itself toward a broader agent economy in its narrative, those currently willing to pay and repurchase are still primarily users already active on-chain.

Fourth, a more concerning signal lies in the ratings. Six out of the ten top products boast a 100% positive review rate, and the rest generally exceed 90%. Notably, CoinAnk maintains a perfect 100% rating despite 16.59K sales—a combination that would be statistically anomalous in any mature rating system. This anomaly directly motivated our deeper on-chain investigation.

The highly concentrated leadership combined with near-perfect ratings naturally raises the question: how much of these figures stems from genuine usage, and how much is artificially inflated by design or human intervention?

To address this question, we conducted further research on several leading products. In pursuit of authenticity, the author personally developed a small-scale product to fully experience the entire process firsthand.

Ultimately, we identified the following key issues.

When analyzing reviews of popular products, two key phenomena emerge:

First, average ratings are universally extremely high—close to perfect scores.

Second, the number of reviews is extremely low, with review rates generally below 1%.

The table below compares the number of reviews against cumulative sales for the top five products by official website review rate, revealing abnormally low review rates.

This is primarily because, under the current product design, usage and reviewing are entirely separate actions. Unless users proactively trigger the review prompt after completing a transaction, the system does not actively remind or guide them—a missing step that directly leads to adverse selection in the review process: only a small minority of users who are either extremely satisfied or extremely dissatisfied bother to return and leave a rating, while the vast majority of ordinary users simply leave after use without ever entering the review flow.

Some agents handle this step even more poorly—not only lacking proactive review prompts, but also omitting the review entry point altogether or offering no review functionality at all. This explains why products listed in the table, despite sales volumes often reaching several thousand or even tens of thousands, have accumulated only a handful of actual reviews, with review rates consistently below 1% to 5%.

Given such a tiny and highly unrepresentative sample, the resulting positive review rate itself warrants serious skepticism.

We selected several top-selling products and conducted a transaction-by-transaction audit of their payment addresses’ on-chain records on X Layer, focusing on three aspects: the frequency of repeated buyer addresses, the temporal distribution pattern of transactions, and the historical activity levels of the buyer addresses themselves.

The audit revealed several consistent patterns:

First, a significant portion of the call records originate from a small number of addresses repeatedly initiating transactions, rather than numerous independent addresses each making a single purchase;

Second, many transactions are highly concentrated along the timeline, exhibiting a pulsed, batch-submission pattern rather than a steady distribution consistent with natural usage rhythms;

Third, a substantial proportion of buyer addresses show almost no other on-chain activity beyond purchasing this service, resembling temporary empty wallets created for a specific purpose rather than long-term active genuine users.

When these characteristics are combined, they diverge significantly from the organic user growth curve typical of a legitimate product and instead closely resemble patterns left by artificial transaction inflation.

Take PixelBrief, currently the top-selling product, as an example: analyzing the latest 20,000 transfer records associated with its Agent address (0xe7BBb197827048BA8fA7E908eC871b80568dBC25) on XLayer yields the following results:

The analysis clearly shows that PixelBrief, ranked first among Agents, corresponds to only 24 unique addresses across its 20,000 sales—with the top ten addresses accounting for 99% of transactions, and the single largest address contributing over half—indicating clear evidence of fake order inflation.

Take the third-ranked 'Kan Dong' (on-chain address: 0x6d60C281f5240aDABf00DD79c9a3Cc37b87D6bf0) as another example: its approach to fabricating transaction volume appears more sophisticated—its nearly 8,000 cumulative sales correspond to 211 unique addresses, far exceeding the top-ranked product’s mere 24 addresses. However, the issue lies in repurchase frequency: a large number of addresses exhibit identical transaction counts.

The table above clearly reveals that among the 211 addresses, 109 executed exactly 50 transfers each, and another 46 executed precisely 47 transfers each. Together, these two groups total 155 addresses (73.5% of all unique addresses) and account for 7,612 transactions, or 97.4% of the total volume.

Going further, examining each group separately:

The 46 addresses in the '47-transaction batch' all completed their activity exclusively on July 26. Sampling individual addresses shows an average active duration of just 4.5 minutes (shortest: 3.8 minutes; longest: 6.2 minutes) to execute all 47 transactions—equivalent to roughly one transaction every 6–7 seconds—after which they immediately went dormant with no subsequent activity.

"Batch 50" involved 109 addresses: primarily concentrated on July 27 (5,315 out of 5,450 transactions). The average active duration per address was 33.5 minutes, again showing a sharp spike followed by a drop to zero.

Meanwhile, an investigation into other top-performing Agent products with abnormal sales volumes revealed that nearly all of the top ten products exhibited similar suspicious patterns of 'few addresses, large transactions,' indicative of potential fake transaction activity.

Of course, it should be noted that on-chain data can only reflect anomalies at the transaction level and cannot directly reveal underlying motives or the parties involved. Our conclusion regarding 'inflated volume' remains speculative; however, the unusually low comment rate provides further supporting evidence for this hypothesis.

The x402 protocol itself addresses only the payment deduction step, but successful deduction does not guarantee synchronous service delivery. Currently, some service providers implement their own callback logic upon integration, and if this implementation is insufficiently robust, users may complete payment successfully while the service endpoint fails to trigger properly, resulting in delayed or missing delivery.

During random testing using a Codex loop script to batch-call a certain Agent, we observed that 25 out of 30 transactions failed. Each attempt completed the x402 signing step, but subsequently failed during the HTTP replay request. Specific causes included:

SSL certificate failures (10 transactions): Python on macOS lacks root certificates by default, causing urllib to raise CERTIFICATE_VERIFY_FAILED when making HTTPS requests. At this point, signing had already succeeded (the 402 response was received prior), but the replay request failed due to the certificate issue.

curl truncation failures (10 transactions): After switching to curl to bypass SSL issues, we found that curl automatically truncates long headers, resulting in an incomplete base64 payload used for signing, which caused the replay request to fail as well.

Timeout failures (5 transactions): The service required OCR plus LLM analysis, taking 60–120 seconds, but the default 60-second timeout caused requests to be terminated mid-process.

Surprisingly, although none of these 30 transactions completed service delivery, all successfully processed X402 payments and were recorded as purchases on the official website.

After fixing the SSL issue and the Curl truncation problem, 10 additional calls were made, of which 8 succeeded and 2 still failed due to timeouts.

In summary:

The SSL issue stems from the user's client-side environment;

The Curl truncation issue arises from a flaw in the platform's protocol design;

Server timeouts result from inadequate deployment mechanisms by Agent developers.

However, the most critical issue is that X402 charges and transfers are being processed before service delivery is completed—this represents the platform's most significant deficiency in Agent vetting at present.

Of course, such issues currently occur more frequently among early-stage service providers who integrated independently rather than using the official Payment SDK, indirectly underscoring the necessity of the official SDK’s standardized signing and replay logic update.

Referring back to the data in Part Five: of the nearly 15,000 tasks posted, more than half never reached completion. A closer examination of these tasks’ actual workflows reveals that the core issue largely lies not with delivery quality, but with earlier matching stages—many tasks, once posted, were never identified or accepted by suitable Agents, remaining stuck in a waiting state until they expired and were closed.

This indicates that the platform’s envisioned narrative of 'Agents autonomously negotiating, bidding, and collaborating' remains largely aspirational: while protocols have been connected, real-world invocation frequency remains limited, and spontaneous collaboration between Agents via the Task Marketplace is still relatively rare.

Raising the issues above is not meant to dismiss OKX.AI itself. When evaluating a product that is just one month old, the standard should not be harsh criticism of its existing flaws, but rather whether its first step holds value and whether its problems are solvable.

First, let’s acknowledge what they’ve done well. Among the similar projects we’ve observed, OKX.AI is one of the few that has managed—within just one month—to operationalize four core components simultaneously: identity (ERC-8004), payments (three parallel tracks: x402, MPP, and a2a-pay), marketplace (Agent Plaza + Task Hall), and arbitration (Evaluator staking-based jury)—with real funds actually flowing on-chain.

Moreover, this system wasn’t built from scratch—it leverages the infrastructure accumulated over years by OKX Wallet, which handles over 1.2 billion daily API calls and supports more than 60 blockchains and 500 decentralized exchanges. This distribution reach and technical foundation is difficult for many point-solution startups to replicate in the short term.

Now, consider the nature of the issues themselves:

Fragmented evaluation mechanisms, on-chain wash trading, occasional payment failures, and low matching efficiency in the Task Hall—these are all typical growing pains of any two-sided marketplace during its cold-start phase. In the early days, when both supply and demand are thin, platforms often tacitly allow—or even facilitate—artificial volume inflation to present more attractive metrics.

Deficiencies in the evaluation system design will likely be addressed gradually as user scale expands and the product iterates.

Within just three weeks, identity- and onboarding-related Skills were consolidated from seven modules down to two; the payment protocol expanded from a single x402 implementation to three parallel tracks; and official SDK support for ASP was added. This pace of iteration demonstrates that the team is rapidly responding to real-world usage issues, rather than abandoning the product after its initial release.

From a broader trend perspective, the significance of OKX.AI may not lie in how perfect it is today, but in its validation of a key insight: the infrastructure required for an agent economy is fundamentally different from the traditional financial infrastructure built for humans.

The rapid emergence of protocols like x402, AP2, and MPP over the past six months shows the industry has moved beyond debating 'whether to build' and into competing over 'whose rails will gain wider adoption.' OKX’s current strategy avoids betting on a single protocol; instead, it integrates all mainstream rails and orchestrates them through its own settlement layer. While this multi-protocol compatibility approach may seem less exciting in the short term than backing a single winning standard, it offers higher fault tolerance.

OKX.AI provides a compelling observational case study for the real-world implementation of Agentic Payment. So, returning to the core question: does this market genuinely exist? And does Agent payment actually hold value?

Our assessment is that value does exist—but it must be understood in two distinct layers.

The first layer is the clear value of micropayments themselves—leveraging on-chain protocols like x402 or MPP to reduce machine-to-machine transaction settlement costs to just a few cents per call, something nearly impossible to achieve with traditional payment networks. Features such as gas-free transactions on X Layer, TEE-managed private keys, and globally reusable identity and reputation systems concretely address the infrastructure gaps that previously hindered the realization of an Agent economy.

The second layer involves the broader narrative of 'autonomous collaboration'—scenarios where Agents automatically negotiate, bid, and deliver services. Current completion data indicates this capability is far from mature. What’s actually operational today are mostly standardized, single-call A2MCP services, not the complex multi-Agent collaborative tasks originally envisioned.

Thus, a more accurate statement might be: OKX.AI has so far demonstrated that the 'machine-to-machine micropayment' leg can already walk, while the 'autonomous Agent collaboration' leg is still in its early, stumbling steps. Bundling both under the same 'Agent economy' narrative risks either overestimating or underestimating one component. What truly warrants ongoing monitoring is whether the task marketplace’s completion rate improves over time and whether the evaluation system gradually sheds its inflated metrics. These two indicators offer more insight than GMV alone into whether this model is genuinely viable.

From a World Cup-themed Agent standing out initially to a gradual expansion across categories; from reliance on a single x402 track to parallel orchestration across three payment protocols—the pace of change on OKX.AI within its first month is, in some ways, more noteworthy than its current transaction volumes.

In truth, it remains far from perfect. Issues like inflated evaluations, inefficiencies in the task marketplace, and occasional fragmentation in the payment flow are real shortcomings that need addressing. Yet, as one of the few platforms to have built a complete, on-chain闭环 integrating identity, payments, marketplace, and arbitration—and actually running it live—it offers a concrete example of how Agent economics might evolve from concept to infrastructure.

We will continue tracking the product’s future iterations and monitor how core protocols like x402, AP2, and MPP progress toward broader real-world adoption.

2042 Views