The Graph says onchain AI agents need identity, structured data and payments to act autonomously.

Artificial intelligence and blockchain have spent years developing largely along parallel tracks. AI has become increasingly capable of reasoning, interpreting instructions and making decisions, while blockchain networks have created programmable financial systems that can operate without centralized intermediaries. The next stage is bringing those capabilities together.

The result could be an internet populated not only by human users, but also by autonomous AI agents capable of discovering information, making decisions, interacting with protocols and paying for services on their own.

That future requires more than increasingly powerful large language models. As The Graph Foundation explains in its Aug. 18 blog post, “The Onchain Agent Infrastructure Stack Explained,” autonomous agents need infrastructure that translates AI reasoning into reliable blockchain actions. The Graph frames the emerging stack around three fundamental requirements: identity, environmental awareness and economic agency.

Identity: Giving agents an onchain passport

Before an autonomous agent can transact, a blockchain needs a reliable way to identify it and determine what it is authorized to do.

Simply giving an AI agent access to a user’s wallet private key creates obvious problems. An incorrect decision or hallucination could expose all of the assets associated with that wallet. It also makes distinguishing between actions performed by a person and those performed by an autonomous system difficult.

ERC-8004, described by The Graph as the Trustless Agents standard, addresses this problem through three onchain registries covering identity, reputation and validation. These registries allow agents to establish recognizable identities and interact without requiring pre-existing trust between participants.

Combined with account abstraction, this model can also create tightly defined permissions. An agent might be authorized to trade only a certain amount each day, for example, without receiving unrestricted control over a user’s funds.

Identity becomes more valuable as agents develop histories. Because actions and feedback can be recorded onchain, agents can accumulate verifiable reputations that other agents and smart contracts can evaluate.

The Graph is supporting this layer through Agent0 Subgraphs, which index agent registrations, metadata, reputation information and validation activity across multiple networks. According to The Graph, this makes it possible for agents to search for other agents by characteristics such as capability or reputation without independently scanning blockchain histories.

Data: Helping AI understand the onchain world

Knowing who an agent is solves only part of the problem. An autonomous agent also needs accurate information about the environment in which it is operating.

Blockchains contain enormous amounts of transparent data, but transparency does not necessarily mean accessibility. Information is distributed across blocks, transactions, events and smart contracts. Asking an LLM to navigate that raw information directly is inefficient and can increase the risk of incorrect conclusions.

This is where blockchain indexing becomes a critical part of agent infrastructure.

The Graph’s Subgraphs organize blockchain information into structured, searchable datasets. When combined with Model Context Protocol, or MCP, that indexed information can become directly usable by AI systems. The Graph describes Subgraph MCP as effectively acting as a translator between agents and complex blockchain data.

Consider an autonomous trading agent tasked with finding an attractive opportunity involving an ETH pair. Before executing a transaction, the agent might need to compare liquidity across protocols, evaluate current conditions and confirm that the opportunity still exists.

Rather than attempting to interpret millions of blockchain logs, the agent could query the relevant Subgraphs through MCP and receive structured information it can reason about.

This illustrates an important distinction in agent infrastructure. The intelligence layer and the data layer solve different problems. An LLM may decide what information it needs and reason about the answer, while indexing infrastructure is responsible for making reliable blockchain information available in a usable format.

Payments: Giving agents economic agency

The final piece is the ability to pay.

Today’s internet payment infrastructure was largely designed around people and businesses. Users create accounts, manage subscriptions, enter payment information or manually authorize transactions. Autonomous software operating continuously cannot depend on those workflows.

The emerging x402 standard offers another model. It revives HTTP’s “402 Payment Required” status code to enable services to request payment directly as part of an internet request.

The Graph has integrated x402 into its Subgraph Gateways, allowing agents to pay for individual queries in USDC without maintaining traditional API accounts or keys. An agent requests data, receives the payment requirement, signs the payment and resubmits the request before receiving the requested information. The Graph’s GraphTally infrastructure handles settlement with Indexers behind the scenes.

The significance goes beyond paying for blockchain queries. Machine-to-machine commerce requires payment systems suited to potentially enormous volumes of small, automated transactions. If agents are constantly purchasing data, computation or services from one another, per-request micropayments can provide an economic model that more closely matches how autonomous software actually operates.

From web3 users to web3 agents

Put these layers together and a clearer picture of the onchain agent stack emerges.

An agent establishes an identity and operates within predetermined permissions. It accesses structured blockchain data to understand current conditions. It can then purchase the information or services it needs and execute an authorized action. The process can repeat without requiring a human to approve every intermediate step.

The Graph calls this process the “agent loop.” ERC-8004 provides identity and accountability, Subgraphs and MCP provide contextual awareness, while x402 and GraphTally support autonomous payments and settlement.

This architecture also points toward a broader shift in how blockchain infrastructure may be designed. Much of web3 today assumes a human is sitting behind a screen, navigating an interface, connecting a wallet and approving transactions. An agent-centric environment requires infrastructure that is machine-readable, programmable and economically autonomous by default.

The Graph already provides blockchain data infrastructure across more than 60 networks and reported that, as of early 2026, it had served more than 1.27 trillion queries to over 75,000 projects. The rise of autonomous agents potentially gives that kind of infrastructure a new class of user: software itself.

AI may provide the reasoning engine for the emerging agentic internet, but intelligence alone cannot create an autonomous economy. Agents also need identities, trustworthy data and native ways to transact. The development of that underlying stack could determine whether onchain AI remains a collection of experiments or becomes a functional machine-to-machine economy operating at internet scale.

Binance has rolled out Agent OS, a new developer platform designed to let AI agents connect to crypto-market data, monitor user accounts, and execute trades on the exchange—subject to permissions and limits set by the user.

In an announcement, the company said the platform supports popular AI tools, including ChatGPT, Claude Code, Codex, and Cursor. Users can authorize agents to view account information and place orders within configured boundaries, while also being able to revoke access and adjust permissions at any time.

Key takeaways

• Binance Agent OS is positioned as a developer platform for AI agents to access exchange data and trade on Binance.
• Authorization controls—permissions, limits, and the ability to revoke access—are central to how agents can act.
• Account separation is supported via dedicated subaccounts, allowing funds and activity to be isolated per agent.
• Exchange monitoring, not agent cognition: Binance can monitor trades placed through Agent OS but cannot see an agent’s external data sources or decision logic outside the chosen AI application.
• Onchain and payment integrations are included, enabling agents to initiate payments and interact with wallets and other onchain services.

What Binance’s Agent OS enables

Agent OS is built around the idea that AI systems should be able to perform structured actions in financial applications—rather than simply providing advice. According to Binance, users can authorize agents to view account information and execute trades on the exchange under a permission model.

The platform is designed to give users practical control over automation. Binance says agents can be assigned to dedicated subaccounts, which can help separate funds and trading activity tied to different agents or strategies. That separation matters for risk management, particularly when multiple agents are running different tasks or operating with different levels of access.

Permissions, limits, and transparency into agent activity

A key detail in Binance’s explanation is what the company can and cannot observe. Binance said it can monitor trades placed through Agent OS, but it does not have visibility into an agent’s external information sources, its interpretation of inputs, or the decision-making process—those occur within the user’s chosen AI application.

This distinction is important for both builders and users. It suggests Binance is implementing guardrails at the exchange-action layer while leaving the reasoning layer to the third-party AI stack. For users, that can reduce exposure to unclear automation behavior, but it also means they still need to carefully audit what their selected AI tools are doing, where they pull information from, and how they translate that information into trading actions.

Binance also emphasized that access is not permanent: users can revoke access at any time and adjust permissions and limits as their needs change.

Beyond trading: payment and onchain connectivity

Agent OS is not limited to market monitoring and order placement. Binance says the platform connects agents to its payment and onchain tools, enabling agents to make payments and interact with wallets and other onchain services.

That broadens the potential use cases for agent automation from trading-centric workflows to wider transaction tasks. For example, an agent might be configured to move assets, execute payments, or coordinate onchain interactions—again within whatever boundaries the user sets.

Binance joins an emerging “agentic” exchange trend

Binance’s move fits into a wider pattern among crypto trading platforms exploring how far AI agents can go in executing tasks. The push is not uniform: different exchanges appear to be testing different levels of autonomy and different product shapes.

In June, Coinbase launched “Coinbase for Agents”, described as a tool that lets AI models such as ChatGPT and Claude connect to user accounts and execute crypto trades and strategies. Coinbase also highlighted support for agent-driven payments via its x402 protocol.

Meanwhile, Kraken reportedly took a more controlled approach in July with an AI-powered investing assistant that monitors markets and recommends trades based on users’ goals and risk preferences, but requires user approval before executing a trade.

Other players have extended the concept beyond direct trading. In a separate development, OKX launched a beta marketplace where AI agents can find work, transact autonomously, and hire other agents for tasks, using stablecoin payments and an onchain reputation system.

The broader narrative has also been reinforced by prominent executives arguing that AI agents may become significant participants in onchain activity. Coinbase CEO Brian Armstrong and Circle CEO Jeremy Allaire have both suggested agents could soon represent a large share of onchain transactions. Binance co-founder Changpeng Zhao has echoed the idea, describing crypto as the “native currency” of AI agents.

What to watch next for Agent OS

With Agent OS, Binance is effectively turning trading permissions into an interface for automation—while keeping the “why” behind decisions inside the user’s AI environment. The next phase for users and developers will likely hinge on how reliably permissions behave in practice, how agents are isolated via subaccounts, and how Binance’s integrations handle real-world onchain and payment flows as more automation moves from demos into production.