ERC‑8004 solves identity, ERC‑8183 solves custody, then what?
Written by: Thejaswini M A
Translated by: Saoirse, Foresight News
For a long period in the development of the internet, search engines only returned a list of results, and all the remaining work was left to the user. What you see is blue titles and hyperlinks. Then we click on the content that looks like it matches our query needs. But when I treat this experience as something in the past, it feels a bit odd—somehow it gives me a sense of aging and time passing.
As a writer, I’ve always been put off by the whole SEO setup. Keywords can ruin a lot of the精彩 content in an article; under Google’s ranking logic, external citations are also seen as votes. In essence, if more other web pages point to yours, your page gets a higher weight. Still, it’s hard to complain. After all, some degree of credibility is better than being completely disregarded. Ads keep the entire internet alive—but evidently, content publishers are very unhappy with this mechanism.
This model is now headed toward extinction. I want to find a book related to finance and economics, or figure out how to handle ankle pain after a long walk—I just open Gemini. Some people use Perplexity, others use GPT. If you’re not sensitive to token consumption, you just choose Claude. Are the results from these tools definitely better than reading an entire article? We don’t know… Maybe not, but it does make life easier. You don’t have to read two articles with completely opposite viewpoints—AI has already helped you deal with the problem.
I think that at least for people under 30, they no longer treat search engines as an entry point that requires proactive access.
Then agents (Agent) appeared, completely breaking the old order. Non-human traffic has surpassed human traffic. An agent executing a research task might read thousands of web pages and ultimately output only a single memo. This is entirely different from humans. Therefore, in an agent-driven economic system, ads may lose their value. I wrote two articles in the past discussing whether software should pay for the resources it takes. Our conclusion was: paying for it is actually the simplest part.
For an agent to complete tasks on the network and settle value, it needs a set of six-layer technology stacks that are independent of each other.
An agent needs to find web resources, trust the returned results, trust other robots, confirm whether the task is completed, complete payment, and also leave appropriate compensation for content creators. Next, we’ll break it down layer by layer in that order.
What’s actually on the real internet?
Both machines and humans use “search,” but their demands from web pages are completely different.
Human search returns a complete list of pages, which users browse by checking titles and choosing where to jump. Machine search for AI agents is different. Since agents are limited by a context window (short-term memory), they do not browse pages completely. They don’t return all links sorted by weight; instead, they extract only a small number of passages most useful for the task and feed them directly to the model to avoid memory overflow. The whole process takes only milliseconds.
The two main companies in this space are Parallel and Exa. Both are independent companies providing agent search services. Parallel was founded by Parag Agarwal after leaving Twitter. It completed a $100 million funding round in April and has a valuation of $2 billion. Parallel executes searches according to clear objectives, returns compressed content snippets. The snippets are ranked by their contribution to the task, and each factual claim includes a source citation and a confidence score.
Now look at Exa: backed by a16z, it completed a $250 million funding round in May at a valuation of $2.2 billion. It serves over 5,000 companies and 40,000 developers, providing underlying capabilities for products such as Cursor, Cognition, HubSpot, OpenRouter, and Monday.com. Exa focuses on semantic understanding rather than keyword matching— even if the original text doesn’t contain the query terms, it can still return matching content. It also returns clean webpage source text directly.
According to Parallel’s official documentation, the cost for 1,000 search requests in the two high-speed tiers is $1. The cost for 1,000 requests in the two low-speed tiers is $5. Regardless of webpage length, URL content extraction is standardized at 1,000 items for $1. Exa charges $7 for 1,000 search requests, which is 7x Parallel. It charges $1 for extracting content from 1,000 pages, and $5 for 1,000 question-answering API calls with citations.
How good are the search results, really?
Google mostly evaluates search quality internally: whether users click the first result, stay on the page, or immediately bounce. It also employs human graders to score the results. There have been tests like RTEC in the past, but there wasn’t an externally public, purchasable leaderboard. Google Analytics and Similarweb measure traffic, not search quality.
In the agent internet, lab-style benchmark test sets have already emerged, specifically measuring whether search tools can help AI obtain the correct answers.
In April 2025, OpenAI released BrowseComp, which includes 1,266 reverse-constructed exam questions. The question setters first determine a verifiable objective fact, then design questions to hide the fact under multiple constraint conditions. If questions are created first, like typical knowledge Q&A, models likely answer correctly based on their own memory or one lucky search. Only by fixing the fact first and then packaging it as questions can we truly test a tool’s web retrieval capability.
Native GPT‑4o scores only 0.6%; GPT‑4o with web browsing capability scores 1.9%; OpenAI’s Deep Research scores 51.5%.
The testing goal is to verify whether the system can find a hidden, verifiable fact in the real internet and output it. All 1266 questions have short, standard answers. The model gets only one chance to answer: a match earns a score, a mismatch earns no score. The final score equals the percentage of questions answered correctly out of the total. Native GPT‑4o’s 0.6% indicates the model can hardly retrieve facts via inherent memory. With browsing enabled, GPT‑4o gets 1.9%, showing the improvement from the basic search tool is minimal. Deep Research gets 51.5%, indicating a longer search‑read loop: roughly half the time it can find the target fact.
Besides BrowseComp, there are other evaluation sets: SimpleQA (OpenAI, 4,326 questions), a factual test set that is easy to grade. As browsing tools sped up, SimpleQA became too easy for models to break, which is why BrowseComp was introduced. GAIA (Meta / Hugging Face, 466 human tasks) targets real assistant scenarios and may require browsing web pages, calling tools, reading files, and multi-step reasoning—not just search. DeepSearchQA is a research dataset for multi-round retrieval; Artificial Analysis performs a comprehensive evaluation based on this dataset. FRAMES (Google DeepMind, 824 samples) is a multi-hop question set that requires extracting facts from multiple sources and then integrating them.
Private institution Artificial Analysis was jointly founded by George Cameron and Micah Hill‑Smith, and published its search index on August 18, 2026. This set of evaluations uses a fixed AI model and evaluation rules for answers; it only changes the search API being called by the agents. If scores change, it means the difference comes from the search tool itself—not from the model’s capability.
Source: @artificialanalysis
Without using search at all, the same model scores 33 on this index baseline and 17 on the BrowseComp subset baseline. Even if the rankings of different search vendors change, the baseline score never changes.
As of August 27, Artificial Analysis’s search index: Perplexity’s mid-tier version ranks first at the 80th percentile; Parallel’s advanced tier and Brave are tied at 75; You.com and Exa are at 74; Firecrawl, Parallel’s base version, and Parallel’s high-speed version are all at 73.
If the search tool’s score is higher than the baseline, it means the search really helps the model obtain new information. The logic is as follows: AI answers based only on memory to get the baseline score. After integrating search, if the score increases, it means the search found information the model originally didn’t know. If the score does not improve, then the search didn’t play a role.
However, the evaluation itself has limitations: the test questions and the models used are not exclusive to any real business scenario. Service providers can optimize for known test sets (e.g., DeepSearchQA) for exam performance, just like websites used to optimize for Google with SEO.
Who is the robot on the other side?
When two AI agents delegate work to each other, having only a wallet address and a profile description is not enough to establish trust. A wallet address can only prove that the account has funds; the text description is easy to forge. It cannot prove the developer identity of the robot, nor can it substantiate whether its past work history is real. The ERC‑8004 standard was designed specifically to solve real identity and reputation verification for agents.
ERC‑8004 was jointly written by people associated with MetaMask, the Ethereum Foundation, Google, and Coinbase. It went live on the Ethereum mainnet on January 29, 2026 and includes three main registries.
First, an identity registry, implemented in the form of NFTs. Each agent corresponds to an numbered token; the token’s uniform resource identifier points to a registry file that records the agent name, service endpoint, whether it supports x402 payments, and which trust models it supports. Token holders are the owners of the agents. When ownership of the token is transferred, the old paying wallet information is cleared, and the new holder must re-sign to prove control.
Second, a reputation registry: it receives signed feedback—i.e., an on-chain public rating system. After using an agent, users can leave ratings on-chain with task tags (e.g., search, rating). The agent owner can’t write reviews for itself; reviews support withdrawal. Anyone can add notes beneath a review—refund records and spam warnings can also be attached alongside the rating. This is a public comment area bound to an agent ID, not a private star-rating system within a specific app.
Third, a verification registry, which connects to third-party independent verification. The agent initiates a verification request, and the verifier’s smart contract returns a 0–100 score along with evidence links. The evidence source can be a task re-run result, zero-knowledge proofs, or proofs produced by a secure chip.
This specification itself does not handle payment logic, and the documentation explicitly states that. It only records an agent’s identity, external evaluations, and third-party verification results; payment is handled by other protocols. The data quality in the registry isn’t ideal. According to the 8004scan July statistics, across 29 public chains there are 385,998 registered agents and more than 460,000 reviews. Most IDs are merely empty badges; 89.2% have not published callable standardized service interfaces. Only 8,631 agents have available services, accounting for just 2.24%.
Source: @8004_scan
In a May study of crawlers targeting Ethereum, BNB Chain, and Base, it was found that on-chain ratings cannot be compared across different agents. Most feedback is not bound to any task result that anyone else can verify, and the cost to spam ratings is very low. After filtering out suspicious reviews, many of the agents that received a large number of ratings no longer have any effective, usable reviews. This standard only unifies the format of rating data output; whether reviews are truly trustworthy is handled by external systems. ENS, EigenLayer, The Graph, and Taiko all say they will adopt this standard, but the rating-filtering tools have not been developed yet.
Did the agents really finish the job?
The Agentic Commerce architecture, jointly designed by the Ethereum Foundation and Virtuals Protocol, aims to enable autonomous AI agents to hire each other, collaborate, and settle on-chain—without any human intermediaries throughout the entire process. ERC‑8004, together with ERC‑8183 (escrow standard), implements this system. The ERC‑8183 standardizes the full workflow between agents: task publishing, fund escrow, result verification, and settlement.
The workflow is as follows: a client agent initiates a task, specifying the task description, deadline, and budget, and designating an evaluator. The client deposits funds into a smart contract. The funds are locked, so neither party can unilaterally use them.
The agent providing the service performs computation off-chain. It uploads the final results to decentralized storage such as IPFS and Arweave; on-chain it only submits a hash or a resource locator. The evaluator (which can be a dedicated judge agent, an automated oracle, or a zero-knowledge proof verifier) checks the delivery.
If the evaluator determines the task is passed, it deducts the service fee and then pays the service provider. If it fails, it refunds the funds to the client. If the service provider goes offline, or if the evaluator stalls beyond the deadline, the contract triggers a public refund logic and returns funds to the client, preventing assets from being permanently locked.
The evaluator can also be directly run by the client itself. The documentation describes this mode as a standard solution in scenarios with no third party. It also clearly specifies: this protocol does not provide dispute resolution or arbitration mechanisms. Rejecting the result / timing out counts as the final determination. If your scoring robot tries to deceive you maliciously, there’s no administrator or customer support to appeal. You have nowhere to go. And regarding the issue mentioned in the previous layer, you also can’t rely on on-chain ratings to select a reliable evaluator—most ratings are fake.
Systems like UMA, Kleros, and Bittensor can mitigate the risk of a single-point evaluator acting maliciously, but they sacrifice speed and increase costs. Pure automated escrow settlement can complete in seconds with extremely low fees. By contrast, upgrading dispute-resolution flows through UMA and Kleros requires challenging windows, posting collateral, and also introducing human participation for voting.
Has the funds transfer been completed?
There’s been a lot of discussion about this recently in the industry. Here’s a brief explanation.
x402 is introduced by Coinbase. It is a layer protocol built on top of existing public chains. It charges per request and uses stablecoins to complete transfers.
Source: @coinbase
Another scheme is the machine payment protocol MPP, jointly released by Stripe and Tempo in March 2026, with Tempo’s mainnet going live on the day of release. MPP is an HTTP standard. The server returns a 402 Payment Required response, along with challenge information. The agent carries payment credentials to retry the request. The server then returns resources and a payment receipt.
Actual fund movement depends on the underlying channel. On the Tempo chain, on-chain settlement for a single transaction takes about 0.5 seconds. Facing a series of small calls, an agent can pre-lock an amount of funds, keep continuous accounting during the process, and then do a single on-chain clearance at the end—rather than initiating an on-chain transaction for every call. Payment limits are determined by the amount locked for this session or controlled by access keys, not a “one-time authorization, permanent deduction” model.
MPP supports multiple underlying channels: Visa issues card specifications, and agents can pay using tokenized cards. Lightspark implements a Lightning Network payment scheme based on the same 402 flow, along with the Visa card program. Merchants already integrated with Stripe card payments can directly receive MPP card payments, with funds entering a normal Stripe account; the Lightning Network is another optional channel under MPP rules. Stripe also supports x402. Both x402 and MPP reuse the HTTP 402 “Payment Required” status code, but they are two different protocols.
Did content creators receive compensation?
If creators don’t get paid, there won’t be new content for the system to search. If creators can’t profit, they’ll stop publishing work, and the AI agents will eventually run out of new reading material.
Cloudflare, Parallel, and OpenLedger are three representative projects in this area.
Cloudflare initially charged a fixed fee for scraping web pages for crawlers. That approach was too rough: scraping the homepage and scraping in-depth investigative reports were charged the same. Now Cloudflare has shifted to a value-based pricing model: it only settles with the publisher when the content truly generates value. Meanwhile, for ad-monetized sites, if an AI service provider refuses to pay, Cloudflare will block AI crawlers.
On May 19, Parallel launched its Index product. Borrowing the Shapley value algorithm from cooperative game theory, it calculates the contribution percentage of each information source to the agent’s final task result, and uses that to determine profit sharing. Instead of paying uniformly by clicks, it measures a single article’s actual contribution to the task through a mathematical model. Cooperative partners include The Atlantic, Fortune, PR Newswire, PitchBook, ZoomInfo, Tracxn, RocketReach, Enigma, and Fiscal AI.
Currently, Index only settles when agents use Parallel’s own search tool. Unless other AI platforms fully integrate, it cannot become a universal standard across the whole network.
OpenLedger is another approach that uses a token system. It measures how much a dataset impacts an AI model, and pays contributors using its own tokens. This scheme is aimed at AI training data developers, not to compensate journalists’ everyday articles.
I tried to find real data on how much these projects actually paid to creators. Cloudflare doesn’t publish payout numbers; Parallel also hasn’t disclosed them; OpenLedger settles with OPEN tokens, and it doesn’t publicly provide an independent summary of the total payouts. Therefore, we currently don’t have solid facts to judge how effective this mechanism is in practice.
In human commercial development history, cash registers actually appeared very late. After humans spent years establishing verbal agreements and local reputation systems, they invented low-friction payments. But the development path of the machine internet is completely the opposite: from the moment it was born, it enabled instant, streaming global settlement. And now the agent economy has to scramble to build on-chain small-claims courts to prevent two robots from defrauding each other over just a few dollars’ worth of tasks.
Finally, we close with a quote from Terry Pratchett:
“To drone on with pure logic—saying that machines merely execute the instructions they receive—is easy to say. But people are entitled to expect that machines have a little bit of common sense, too.”