Agentic AI in Crypto: An Evergreen Explainer

Unlike traditional chatbots that simply answer questions, agentic AI describes systems that can understand a goal, plan multi-step actions, and then execute those actions autonomously across software and financial rails. In crypto, that increasingly means AI agents that can hold keys, move assets, and transact onchain in stablecoins like USDC with minimal human input, while still operating under clear guardrails and user control.

What Is Agentic AI?

At its core, agentic AI is about giving software agency: the ability not just to generate text or predictions, but to interpret objectives, formulate plans, and act in the world through tools, APIs, and transactions. Whereas a conventional large language model will respond to a prompt with an answer, an agentic system might take a high-level instruction such as “find me the best hotel in Bangkok under a hundred dollars” and autonomously decide which services to call, what parameters to use, how to compare options, and when to come back for human approval. The International Monetary Fund describes these “agentic” systems as AI that can interpret objectives, plan multistep actions, and interact with digital services with limited human intervention, highlighting the step change from static automation scripts to flexible, goal-driven agents. Academic and industry discussions increasingly converge on this notion of goal-directed, tool-using autonomy as the defining feature of agentic AI, even as individual implementations differ in how much control they delegate to the machine.

MIT’s framing adds another important dimension: coordination across multiple agents. Rather than relying on a single, monolithic AI that tries to do everything, agentic architectures often orchestrate specialized agents that collaborate on parts of a workflow, much like microservices in modern cloud applications. One agent might specialize in retrieval, another in planning, another in calling financial APIs, and a coordination layer routes tasks and context between them. This multi-agent pattern is particularly relevant for crypto, where different agents may need to interface with price oracles, smart contracts, custodial or non-custodial wallets, compliance services, and user interfaces while maintaining a clear chain of responsibility.

The rise of agentic AI coincides with the broader maturation of generative models and tool-use protocols. Industry analysts expect the agentic AI market to grow from around 7.8 billion dollars today to more than 52 billion dollars by 2030, and Gartner forecasts that roughly forty percent of enterprise applications will embed AI agents by the end of 2026, up from less than five percent in 2025. This shift is not just a matter of plugging more models into software. It reflects a structural evolution in how systems are designed, moving from passive assistants and static “if-this-then-that” automation toward persistent, context-aware entities that keep working on your behalf even when you are offline.

From Chatbots to Autonomous Agents

A useful way to understand agentic AI is to contrast it with the first wave of chatbot-style generative AI. Traditional chat interfaces are essentially stateless or lightly stateful interactions: the model processes a prompt, consults its context window, and returns an answer. If you ask it to book a flight or rebalance a portfolio, it can explain how to do so, but without additional scaffolding it cannot safely log into accounts, read live data, or execute transactions. The user remains the primary actor; the AI is an advisor.

Agentic systems insert an additional layer between the model and the environment: an execution harness that can call external tools, handle multi-step workflows, and maintain long-term memory for a specific user or task. This is visible in consumer-facing frameworks like Hermes agents, which now offer session recall so an agent can remember what you worked on in previous sessions, run background tasks that continue while you do other things, and even control your computer to click, type, and navigate on your behalf. Hermes agents can also spin up coding environments through tools like Codex CLI, search real-time social media data via OAuth-based integrations, and generate video directly from text, all orchestrated within a continuous agentic loop rather than a series of isolated prompts.

The result is that users start thinking of these systems less as chatbots and more as digital colleagues or “junior executives.” In crypto contexts, that shift is significant. A trading agent that can remember your risk preferences, monitor markets, and execute predefined strategies while respecting onchain limits is qualitatively different from a chatbot that merely suggests what you might trade. Likewise, a payments agent that can move USDC across chains based on rules you set effectively becomes a programmable financial entity in its own right, one that must be designed and governed with the same care as any other economic actor.

Key Characteristics of Agentic AI

While implementations vary, most agentic AI systems share several technical and behavioral characteristics. First, they are goal-driven rather than request-driven: you specify an outcome (“optimize my travel budget over the next quarter” or “keep my portfolio delta-neutral within a given volatility band”), and the agent decomposes that outcome into concrete steps. Second, they are tool-using. Instead of only generating natural language, they call APIs, invoke code, query databases, and interact with devices to carry out those steps in the real world. Hermes’ ability to control a user’s computer, for example, illustrates how agents can extend beyond pure text interfaces and manipulate arbitrary user interfaces to complete tasks.

Third, agentic systems typically maintain memory over time. Rather than treating each request in isolation, they store structured information about user preferences, past decisions, and environmental state so that future actions can be personalized and consistent. Enhancements like session recall in Hermes show how agents are increasingly able to recall past workstreams and build on them, moving closer to the notion of persistent digital personas. Fourth, advanced frameworks orchestrate multiple cooperating agents, each optimized for a particular role—planning, research, execution, or monitoring—coordinated by an overarching controller that decides which agent does what, and when.

Finally, there is a growing emphasis on AgentOps, a discipline focused on monitoring, evaluating, and governing highly autonomous systems after deployment. Researchers in software engineering stress that agentic systems require continuous observability, intervention capabilities, and oversight, not just static testing prior to launch. This is especially crucial in crypto, where misconfigured or malicious agents could move real value at machine speed, exploit arbitrage opportunities in unintended ways, or interact with insecure smart contracts. AgentOps practices—ranging from policy enforcement layers to kill switches and anomaly detection—are becoming as essential to agentic AI as DevOps is to traditional software.

Multi-Agent Architectures and Microservices Thinking

Another defining trend is the shift from single, all-purpose agents to networks of specialized agents, mirroring the microservices revolution in conventional software architecture. Machine learning commentators describe the field as moving away from monolithic “do-everything” agents toward orchestrated teams that can be composed and recomposed depending on the task. In practice, this might mean a research agent that gathers and summarizes data, a planning agent that chooses a strategy, an execution agent that interfaces with external APIs, and a monitoring agent that watches for anomalies—all coordinated via a protocol like Anthropic’s Model Context Protocol or Google’s Agent-to-Agent Protocol, which aim to be HTTP-like standards for agent interaction.

In crypto and finance, this multi-agent pattern aligns naturally with the modular structure of onchain systems. A DeFi-focused agentic stack might include one agent specializing in DEXs and liquidity pools, another in centralized venue routing, another in compliance checks such as travel rule screening, and yet another in analytics and reporting. A coordination layer then decides when to route orders to an onchain AMM versus an order-book-based centralized exchange, taking into account gas costs, slippage, and market impact. The orchestration logic itself can live offchain, but its decisions can be audited by logging signed messages or proofs onchain, providing an accountability trail that fits the expectations of both regulators and sophisticated users.

Danicjade

Apr 24, 2026

View article →

UAE unveils two-year roadmap to embed agentic AI across 50% of government operations, positioning AI as an “executive partner” in public sector workflows

crypto.news • Apr 24, 2026

Top Comment

Benthic

Apr 24, 2026

G42's Presight already runs automated workflows across multiple UAE ministries — 50% is as much rebadging as new deployment. The crypto-relevant test: does VARA licensing inherit the agentic treatment? VASP applications clearing in weeks vs months changes UAE's domicile pull against Singapore and Hong Kong.

Why Crypto Cares About Agentic AI

For crypto, agentic AI is not merely another integration story; it touches the core idea of autonomous, programmable value. Blockchains already enable smart contracts that execute logic deterministically once conditions are met. Agentic AI extends this logic beyond predefined code paths into the realm of open-ended reasoning, decision-making, and adaptation. When an AI agent can read market conditions, negotiate with other agents, and move USDC or other assets onchain, it becomes a participant in crypto’s economic systems rather than just an advisor on the sidelines.

Crypto’s emphasis on transparent, tamper-resistant ledgers also offers something the wider AI ecosystem badly needs: trust and verifiability. If an agent makes a series of payments, trades, or governance decisions, anchoring those events onchain provides an immutable record that can be audited after the fact. That record can support legal accountability, performance analysis, and risk controls in ways that purely offchain logs cannot. As AI systems gain more freedom to act, the ability to prove what they did, when, and under which constraints becomes invaluable.

Aligning Autonomous Software with Autonomous Value

The alignment between agentic AI and crypto becomes even clearer when you consider machine-to-machine payments and onchain machine economies. Projects like Aptos explicitly highlight “autonomous AI trading systems” and institutional markets as key areas for future onchain activity. The Aptos Foundation and Aptos Labs have committed around fifty million dollars across their stack to fund products, research, protocol infrastructure, and a dedicated fund for trading and AI partners, emphasizing that institutional desks need deep order books, MEV protection, and connectivity to existing systems before they can route significant volumes onchain. Those prerequisites are exactly what agentic trading systems will rely on as they become more prevalent.

Similarly, infrastructure projects like Orbs are repositioning themselves around agentic AI and onchain trading. Orbs describes itself as a decentralized Layer-3 blockchain focused on advanced onchain trading, and its V5 upgrade—implemented via a Committee Sync MVP across Ethereum and Arbitrum—is designed to improve how decentralized trading execution is verified across chains while strengthening infrastructure for agentic AI and crypto trading applications. In effect, Orbs aims to become a specialized execution and verification layer where autonomous trading agents can route orders, have them validated by a decentralized committee, and anchor proofs of correct execution back to base chains.

These examples illustrate a broader convergence: as more value flows into programmable assets and decentralized markets, there is growing demand for systems that can monitor, optimize, and act automatically on that value. Agentic AI provides the cognitive and decision-making layer, while crypto provides the settlement and verification layer. Together they enable a spectrum of agentic participants, from simple rebalancing bots to sophisticated cross-chain arbitrageurs and portfolio managers that can respond to markets in real time.

Crypto Primitives as Enablers of Economic Agency

Crypto also offers a set of primitives that map neatly onto the needs of agentic AI. Non-custodial wallets and smart contracts give agents programmable access to funds under strict, auditably enforced rules. Stablecoins like USDC provide a relatively predictable unit of account that is easy for an AI to reason about, especially when denominating budgets, limits, and risk thresholds. Protocols like the x402 standard, which Travala uses in its travel protocol, enable instant stablecoin payments directly over the internet for APIs, apps, and AI agents without manual checkout flows. By abstracting away traditional card networks and consumer checkout frictions, x402 allows an agent to settle payments directly with services over the web, often with gasless USDC transactions on networks like Base.

The importance of these features becomes obvious in use cases like autonomous travel booking. Travala has launched what it calls the world’s first end-to-end agentic AI travel protocol, enabling autonomous agents to search, book, and pay for stays at more than 2.2 million hotels—including major chains like Marriott, Hilton, and IHG—without human involvement until the final payment authorization step. The Travala Travel Model Context Protocol (MCP) runs on Base and integrates x402 so that agents can execute instant, gasless USDC payments directly from within a chat or agent environment. Developers who integrate their agents with this MCP even receive a programmatic ten percent rebate in Coinbase’s wrapped Bitcoin (cbBTC) for every successful booking, settled onchain to their wallets, effectively creating a native revenue stream for agentic commerce integrations.

This pattern—agents that can move stablecoins over open payment standards to transact with APIs and earn protocol-native rewards for doing so—offers a blueprint for how crypto and agentic AI can reinforce each other. Crypto provides the programmable, composable money; agentic AI provides the intelligence that decides how, when, and why that money moves.

The Agentic Web and Web4

Some builders view this convergence as a step toward an “agentic web” or Web4, where the primary actors online are not human users manually clicking links but autonomous AI agents acting on their behalf. Animoca Brands, for example, describes the agentic web as a decentralized ecosystem of persistent AI agents with memory, able to negotiate, collaborate, and transact independently for their human stewards, executing complex intents rather than merely providing information. Its Minds platform is pitched as a persistent AI agent layer that lets users deploy sovereign, always-on agents—called “Minds”—without running local servers or managing hardware.

Minds emphasizes control and customization, aiming to remove complexity while preserving full sovereignty for both builders and end users. The platform is also backed by a ten million dollar investment program for projects that make Minds a core product layer, with Animoca Brands positioning this as foundational infrastructure for the agentic web. In this vision, users might maintain multiple Minds—one focused on DeFi strategy, another on gaming assets, another on social and productivity—each interacting directly with protocols, marketplaces, and other agents. Blockchains become the ledger where these agents’ identities, reputations, and assets live, while agentic AI provides the behavioral layer that animates those identities into economically active entities.

Seen through this lens, crypto’s role is not simply to power a few agentic features but to anchor a new phase of the web where autonomous software entities interact through open, programmable value rails.

How Agentic AI Systems Work

Although implementations differ, most agentic AI stacks share a similar high-level architecture consisting of a planning core, tool integrations, memory systems, and an oversight layer. Understanding these components helps clarify how agents move from a user’s plain-language intent to concrete actions onchain or in traditional financial systems.

Architecture: Orchestrators, Tools, and Memory

At the center of an agentic system sits an orchestrator, often implemented as a large language model or a collection of models that interpret user instructions and decompose them into steps. The orchestrator is responsible for deciding which tools to call, how to handle intermediate results, and when to ask the user for clarification or confirmation. Tool integrations expose specific capabilities—such as querying a hotel inventory API, sending a USDC payment over x402, placing a trade through a DeFi protocol, or reading from an onchain data indexer—to the agentic core in a format it can reason about.

Memory plays a crucial role in making these systems feel coherent and trustworthy. Short-term or “working” memory keeps track of the current task, while long-term memory stores stable facts about the user, preferences, and past interactions. Upgrades like Hermes’ session recall illustrate how memory can become more granular, allowing an agent to answer questions such as “what did we work on last Thursday?” or resume multi-day projects without re-prompting. In a crypto context, memory might also include historical portfolio holdings, realized gains and losses, risk tolerances, and blacklisted assets or addresses, enabling agents to act consistently over time.

The final piece is an oversight and control layer that enforces policies. This layer can implement hard limits on spending, define which smart contracts or domains the agent may interact with, require explicit user approval for certain thresholds or actions, and log all activity for audit. In banking-style deployments such as Payman AI, which deploys agentic AI to execute real banking transactions—payments, transfers, and account analysis—over existing rails via voice or text, this layer is central to ensuring that agents are “under your control” even as they automate core operations.

Standards and Protocols: MCP, A2A, and x402

As agentic AI systems become more complex, standardized protocols are emerging to make tools and agents interoperable. Anthropic’s Model Context Protocol (MCP), for example, provides a structured way for agents and tools to exchange context and capabilities, making it easier to plug new services into an agentic environment. Travala’s Travel MCP uses this type of protocol to allow agents to search, book, and pay for travel within a single conversational or agentic interface, while Coinbase’s Agentic Wallet MCP acts as a connector that lets agents initiate and sign crypto payments. Developers integrating with Travala’s Travel MCP can configure travel booking “skills” for agents and connect via remote MCP servers, enabling a smooth path from user intent to onchain settlement.

Google’s Agent-to-Agent Protocol (A2A) aims to play a similar role for agent coordination, setting standards for how agents discover each other, negotiate responsibilities, and exchange data. Machine learning practitioners increasingly describe MCP and A2A as analogous to HTTP for the agentic era, foundational protocols that sit beneath a wide variety of applications and services. Their importance for crypto is straightforward: an agent that can speak MCP or A2A and understands how to use an x402 payment endpoint, a Base RPC endpoint, or a DeFi protocol API can interoperate across a wide range of chains and services without bespoke, brittle integrations.

The x402 protocol itself is a key example of how payment standards are being designed with agents in mind. Travala uses x402 as an open payments standard that enables instant stablecoin payments directly over the internet for APIs, apps, and AI agents, removing the need for manual checkout flows. When combined with gasless USDC transactions on a network like Base, x402 lets an agent settle transactions in the background, only surfacing crucial checkpoints—such as final authorization—to the user. This is precisely the kind of pattern the IMF highlights when arguing that agentic AI will reshape payments: AI systems interpreting objectives, planning multistep actions, and interacting with digital services in ways that blur the line between user instruction and automated execution.

AgentOps: Monitoring, Evaluation, and Guardrails

Once an agentic system is deployed into production, the challenge shifts from building it to operating it safely. Researchers convening at venues like the Agentic Engineering workshop at ICSE emphasize that highly autonomous systems require continuous monitoring, evaluation, observability, intervention, and oversight—collectively described as AgentOps. Unlike traditional software, where behavior is largely fixed by code and configurations, agentic systems can exhibit emergent behaviors as models, data, and external conditions change.

In practice, AgentOps frameworks track metrics such as task success rates, error types, tool usage patterns, latency, and cost, while also surfacing anomalous behaviors for human review. For crypto-related agents, additional metrics include realized and unrealized P&L, slippage relative to benchmarks, adherence to risk limits, and compliance with region-specific policies. Some organizations are designing “enterprise agentic automation” frameworks that combine dynamic AI execution with deterministic guardrails and human judgment at key decision points, rather than relinquishing full control to opaque models.

Cost optimization is becoming a first-class concern in AgentOps, similar to how cloud cost management emerged as a discipline in the microservices era. As agents call multiple tools, use long-context models, and interact with complex workflows, computational and transaction costs can escalate quickly. Businesses adopting agentic AI at scale—whether for trading, support, or commerce—must therefore manage not only performance and safety but also economic efficiency, especially when agentic systems are making frequent onchain calls and paying gas or protocol fees.

Squidalik

Mar 18, 2026

View article →

By 2029, agentic AI could handle 80% of customer service tasks, cutting costs by 30%

𝕏/@Fetch_ai • Mar 18, 2026

Top Comment

Danicjade

Mar 19, 2026

Cutting down costs thereby cutting down jobs. People have to prepare for what is to come

Real-World Crypto Use Cases for Agentic AI

The theoretical appeal of agentic AI in crypto is already translating into production systems across travel, commerce, trading, payments, and even public-sector operations. These examples help clarify both the promise and the constraints of letting agents handle real money.

Travel and Agentic Commerce

Travala’s agentic travel protocol is perhaps the clearest demonstration of end-to-end agentic commerce in a crypto-native setting. By launching what it bills as the world’s first end-to-end agentic AI travel protocol, Travala enables autonomous agents to handle the full booking journey—searching inventory across more than 2.2 million hotels, comparing options, and initiating payments—before the user steps in to confirm and authorize the final charge. The system uses a dedicated Travel MCP that agents can connect to, and it is built on Base, Coinbase’s Layer-2 blockchain, which offers low-cost, high-throughput settlement optimized for this kind of transactional workload.

The payment layer relies on USDC and the x402 protocol, which allows instant, gasless stablecoin payments directly between APIs, applications, and AI agents. From the user’s perspective, this means that an AI agent can, for example, interpret a natural-language instruction like “book me the best hotel in Bangkok under a hundred dollars,” find a suitable option, and then pay using USDC on Base without the user ever touching a traditional checkout page. For developers, Travala offers a rebate program that sends ten percent cbBTC back to their wallets for each successful booking facilitated by their agents, creating an onchain revenue stream that aligns incentives for building agentic integrations.

The broader retail sector is exploring similar patterns of agentic commerce. Rezolve AI, for instance, has partnered with Tata Consultancy Services (TCS) to scale agentic commerce globally through its intelligent commerce platform, brainpowa. Under this partnership, TCS will help retail enterprises embed agentic AI into core commerce workflows, enabling AI-led experiences across conversational commerce, intelligent discovery, and agentic checkout. Although Rezolve’s work is not inherently crypto-native, the underlying pattern—agents handling discovery, recommendation, and checkout—aligns closely with what we see in crypto-centric environments like Travala, especially as more merchants adopt stablecoins and onchain loyalty or rebate schemes.

Trading, DeFi, and Onchain Machine Economies

In trading and DeFi, agentic AI is emerging as a natural extension of the automated bots and quant strategies that already dominate volumes. Projects such as Orbs are explicitly positioning their infrastructure as a foundation for agentic trading. Orbs’ V5 upgrade, centered on a Committee Sync MVP deployed on Ethereum and Arbitrum, is designed to improve cross-chain verification of decentralized trading execution and to strengthen infrastructure for agentic AI and crypto trading applications. By providing a Layer-3 environment tailored to advanced trading, Orbs offers agents a specialized domain where they can route orders, rely on decentralized committees to validate execution, and anchor proofs back to underlying chains.

Aptos’ ecosystem strategy underscores how Layer-1s are thinking about agentic AI in trading. The Aptos Foundation and Aptos Labs have committed around fifty million dollars to expand their tech stack and fund ecosystem projects, focusing on institutional markets and autonomous AI systems that transact onchain. This capital is directed toward products, research, protocol infrastructure, and a fund for external trading firms and AI teams building on Aptos. The stated rationale is that institutional desks require robust order-book depth, MEV protection, and seamless connectivity to existing systems before they can route large flows to onchain venues, especially when those flows may be partially or fully controlled by agentic systems.

On the buy side, hedge funds are already experimenting with agentic AI for data management and decision support. Platforms like Unique allow funds to deploy autonomous agents that can ingest vast quantities of unstructured data, synthesize key insights, and surface summaries tailored to specific workflows, turning what used to require hours of manual scanning into digestible, actionable intelligence. These agents operate within secure, flexible AI infrastructures that let funds maintain control over data and models while automating repetitive cognitive tasks. While many funds still keep the final trading decision under human control, the trajectory is clearly toward agents taking on more of the research, monitoring, and execution pipeline.

Within the exchange sector, some commentators argue that agentic AI will be a competitive necessity rather than a novelty. Analysts such as Joules at Sahara AI contend that crypto exchanges, which historically differentiated themselves on liquidity, fees, and listings, are now converging on similar feature sets and must adopt agentic AI or risk failure. The argument is that exchanges will need agent-driven features—such as personalized trading copilots, automated risk controls, and smart routing agents that work on the user’s behalf—to stand out and serve both retail and professional users as markets become more automated.

Platforms like DeAgentAI extend this logic by offering infrastructure for deploying autonomous AI agents with identity, continuity, and consensus on chains such as Sui, BNB Smart Chain, and Bitcoin. By providing a standardized way to give agents onchain identities and have their decisions validated or mediated by consensus mechanisms, such platforms aim to make agentic trading and machine economies more robust, auditable, and interoperable.

Payments, Banking, and Stablecoins

Agentic AI is also reshaping the frontiers of payments and banking, where conversational interfaces and autonomous workflows intersect with highly regulated money flows. The IMF has argued that agentic AI systems capable of interpreting user objectives, planning multi-step actions, and interacting with digital services will have profound implications for how payments are initiated, routed, and reconciled. In an agentic environment, users may no longer manually initiate individual transactions; instead, they set policies and budgets, and agents manage the details.

Startups like Payman AI exemplify this approach in the traditional banking sector. Payman advertises “agentic AI that does the banking,” deploying systems that can execute real banking transactions—payments, transfers, and account analysis—via voice or text, running on customers’ existing banking rails. The idea is that customers can speak or type instructions such as “pay the electricity bill from my checking account” or “transfer five hundred dollars to savings when my paycheck lands,” and the agent handles authentication, scheduling, and execution. While Payman’s focus is on bank accounts, the same pattern maps cleanly to stablecoin wallets and onchain accounts managed through MPC or hardware-secured keys.

On the crypto-native side, the x402 protocol and USDC-based flows in architectures like Travala’s Travel MCP illustrate how payments can be designed from the ground up for agentic interaction. Because x402 is an open standard that allows APIs, apps, and agents to receive instant stablecoin payments over the internet without manual checkout flows, it is well suited to environments where agents must frequently settle microtransactions, pay for data, or trigger conditional payments in response to events. When combined with gasless USDC transactions on Base, this yields a user experience where AI agents can pay as they go while the user maintains high-level control over budgets and approvals.

Public Sector and Government Operations

Governments are also experimenting with agentic AI, sometimes in ways that intersect with digital identity and payments. The United Arab Emirates, for example, has announced a two-year target to convert half of all federal government operations, procedures, and services to agentic AI models. In public statements, officials frame AI as an “executive partner” in government workflows, implying that agentic systems will take on operational responsibilities that go beyond simple chat-based interfaces for citizen support.

For crypto, such initiatives raise intriguing possibilities and hard questions. If government services—from licensing to social benefits—are mediated by agentic AI, integrating with national identity systems and potentially with central bank digital currencies or regulated stablecoins, the line between AI, finance, and governance becomes increasingly blurred. Agentic systems might eventually interact directly with onchain identities and assets for tasks such as tax collection, compliance reporting, and benefits disbursement. At the same time, the stakes for robustness, transparency, and oversight become extremely high, particularly when agents act at scale on behalf of the state.

Everyday Productivity and Consumer Agents

Finally, consumer-facing agent frameworks such as Hermes show how agentic AI is filtering into everyday workflows, laying the behavioral groundwork for more financially capable agents down the line. Hermes’ recent updates include enhanced memory via session recall, allowing agents to remember everything done in prior sessions and answer questions about past activities, as well as background task execution that lets users queue up multiple tasks for the agent to work on asynchronously. Integrations with Grok via OAuth enable real-time search of social media posts, while native Codex CLI support lets agents autonomously spin up coding sessions for development tasks.

Hermes agents can also control a user’s computer, seeing the screen, clicking, and completing tasks, effectively acting as remote RPA (robotic process automation) bots guided by language-level instructions. Additional features such as native video generation and auto-generation of Kanban tasks, where high-level goals are decomposed into subtasks and assigned to sub-agents, further illustrate how agentic architectures are shifting toward multi-agent coordination and complex, ongoing workflows. Although these use cases are not inherently crypto-specific, they are directly applicable to crypto once agents can safely hold keys and access wallets. The same agent that today manages your calendar and drafts emails could tomorrow monitor your positions, claim staking rewards, and rebalance your portfolio within constraints you define.

Infrastructure: Chains, Data, and Compute for Agentic AI

The rise of agentic AI is putting new demands on crypto infrastructure, from base layers and rollups to specialized execution environments and data services. Agents need low-latency, low-cost, and highly available rails to move value; they also need reliable identity and state anchoring, as well as robust pathways to offchain data and compute.

Layers 1, 2, and 3: Where Agents Settle and Execute

Different blockchain layers are positioning themselves as natural homes for agentic workloads. Base, the Ethereum Layer-2 developed with Coinbase, provides the settlement layer for Travala’s agentic travel protocol, taking advantage of Base’s scalability and low-fee environment to support high volumes of small, USDC-denominated payments. Because Base inherits Ethereum’s security while offering cheaper transactions, it is well-suited for agent-driven commerce where agents may need to make frequent payments on behalf of users.

Aptos, by contrast, is a high-throughput Layer-1 that is explicitly investing in infrastructure tailored to institutional flows and agentic AI workloads. The fifty million dollar commitment from the Aptos Foundation and Aptos Labs is intended to expand the chain’s tech stack for trading and AI workloads, fund ecosystem products and research, and establish a fund for trading firms and AI teams building on Aptos. This suggests a strategy where the base layer itself is optimized to support high-volume, low-latency trading traffic, including flows initiated or managed by autonomous agents.

Orbs represents yet another approach, positioning itself as a Layer-3 that sits atop Layer-1 and Layer-2 chains to provide specialized infrastructure for advanced onchain trading and agentic AI. Its V5 Committee Sync upgrade, deployed first on Ethereum and Arbitrum, is designed to coordinate a decentralized committee that verifies trading execution across chains, bolstering trust in agent-driven trades. By offloading complex logic and cross-chain coordination to a dedicated layer, Orbs aims to make it easier for agents to execute sophisticated strategies without burdening the base layers with additional complexity.

Platforms like DeAgentAI, which runs AI infrastructure across chains like Sui, BNB Smart Chain, and Bitcoin, further illustrate the multi-chain reality of agentic AI. Rather than binding agents to a single chain, DeAgentAI offers tools to deploy autonomous agents with onchain identity, continuity, and consensus primitives across multiple networks, allowing them to participate in diverse ecosystems while maintaining coherent state and governance.

Identity, Data, and Sovereignty for Agents

Agentic AI also forces a rethinking of identity and sovereignty. Animoca’s Minds platform, for example, emphasizes “sovereign, always-on AI agents” that users can deploy and direct without managing local infrastructure. Each Mind can be customized and controlled by its human steward, with persistent memory and the ability to negotiate, collaborate, and transact independently within the agentic web. In this model, agents resemble autonomous entities with their own identities and capabilities, albeit constrained by human-defined policies.

Blockchains provide a natural substrate for anchoring these identities and for tracking agents’ reputations, permissions, and histories. An agent’s onchain address or DID can serve as its canonical identifier, while smart contracts can encode what that agent is authorized to do—how much it can spend, which protocols it can access, which other agents it can delegate to, and how it is governed or upgraded. Platforms like DeAgentAI reinforce this by offering identity, continuity, and consensus layers specifically designed for autonomous AI agents, effectively turning agents into first-class onchain citizens.

Data access is equally critical. Agents need reliable access to onchain and offchain data, from price feeds and order books to user-specific documents and transaction histories. This is driving demand for data infrastructures that can serve AI agents efficiently and securely. Hedge funds using Unique’s platform, for instance, rely on agentic systems to ingest and synthesize large volumes of data into actionable insights, which implies robust, scalable data pipelines and governance. For crypto, oracles, subgraphs, and indexers will need to become AI-friendly, providing interfaces that agents can query and reason about with minimal friction.

Tooling, Ecosystems, and Funding

The agentic AI wave is also reshaping how ecosystems nurture developers and startups. Animoca Brands’ ten million dollar investment program for projects building on the Minds platform is explicitly aimed at teams that are serious about agentic AI and make Minds a core product layer. The associated “Build East” demo day, co-hosted with Hong Kong Science and Technology Parks Corporation (HKSTP), provides a fast-track evaluation path for local agentic AI teams, with selected projects gaining potential access to investment, developer support, and introductions across Animoca’s portfolio of more than six hundred Web3 companies.

These efforts are part of a broader three-tier ecosystem forming around agentic AI. Analysts describe Tier 1 as hyperscalers providing foundational infrastructure such as compute and base models, Tier 2 as established enterprise software vendors embedding agents into existing platforms, and Tier 3 as “agent-native” startups building products with agent-first architectures from the ground up. Crypto projects sit uneasily across these tiers. Some chains and L2s function like Tier 1 infrastructure for onchain settlement; wallets, exchanges, and DeFi protocols resemble Tier 2 platforms embedding agents into their interfaces; while new agent-native dapps and protocols fall into Tier 3.

The Hermes ecosystem shows how community-driven tooling can accelerate adoption. By giving users and developers a flexible agent framework with features like background tasks, computer control, coding, and content generation, Hermes makes agentic capabilities accessible to a broad audience, which in turn creates demand for deeper integrations—including with crypto rails and wallets. Over time, we can expect similar frameworks to emerge that are explicitly crypto-native, offering out-of-the-box support for signing transactions, interacting with DeFi protocols, and coordinating across multiple chains and L2s.

JLJohn

Mar 17, 2026

View article →

Lyft to make rides smarter and more efficient through Agentic AI, accelerate AV future with NVIDIA DRIVE Hyperion.

investor.lyft • Mar 17, 2026

Top Comment

Danicjade

Mar 17, 2026

So much NVIDIA news today, they are really working hard in this AI sector

Risks, Regulation, and Open Questions

As agentic AI systems gain access to real money and financial infrastructure, the risks and regulatory questions they raise become more urgent. Crypto’s existing challenges around security, market integrity, and compliance are magnified when agents rather than humans are the ones clicking buttons and signing transactions.

Technical and Operational Risk

On the technical side, agentic systems can fail in ways that are both subtle and catastrophic. Mis-specified goals, ambiguous instructions, or bugs in tool integrations can lead agents to take actions that are harmful or unintended, especially when those actions involve moving funds or adjusting leveraged positions. Because agents can operate continuously and at machine speed, even a small misconfiguration can result in large losses before humans notice.

AgentOps frameworks aim to mitigate these dangers by providing continuous monitoring, evaluation, and intervention capabilities. However, building robust AgentOps for crypto-facing agents is especially challenging, because the consequences of failure are not just degraded user experience but actual financial loss and potential systemic impact. The need for deterministic guardrails and human oversight at critical junctures is widely recognized in enterprise deployments, where organizations design systems that combine dynamic AI execution with deterministic checks and approval workflows. In crypto trading and DeFi, similar patterns will be essential, such as requiring explicit human approval for large trades, withdrawals, or changes to strategy parameters.

Key management is another critical concern. Giving agents the ability to sign transactions implies either storing private keys in environments the agent can access or using MPC and hardware-enforced abstractions to allow limited signing without exposing raw keys. Any compromise in these mechanisms—whether through prompt injection, tool-chain vulnerabilities, or underlying infrastructure breaches—can result in irreversible loss. Agentic systems interacting with non-upgradeable smart contracts must be especially cautious, since mistakes cannot easily be rolled back.

Market Integrity and Systemic Risk

From a market integrity standpoint, the proliferation of agentic trading and liquidity management systems could amplify existing concerns around algorithmic trading and high-frequency strategies. If many agents adopt similar models or respond to similar signals, their actions could become correlated, potentially exacerbating volatility or triggering feedback loops during market stress. The IMF’s discussion of agentic AI in payments hints at broader macroprudential concerns when AI systems coordinate large flows of money based on learned patterns rather than explicit rules.

Infrastructure projects like Orbs and Aptos are aware of these issues and emphasize features such as order-book depth, MEV protection, and cross-chain execution verification as prerequisites for institutional flows and agentic participation. By offering mechanisms to detect and mitigate front-running, ensure fair ordering, and audit execution across chains, they aim to provide a safer environment for agentic trading strategies to operate at scale. However, the potential for new forms of market manipulation or emergent behaviors remains an open question, particularly when agents interact across multiple venues and jurisdictions.

Regulation, Accountability, and Governance

Regulators face a complex challenge in determining how to treat agentic AI systems that handle financial tasks. When an AI agent misroutes funds, executes an unauthorized trade, or fails to comply with sanctions screening, who is responsible? Is it the user who configured the agent, the platform that deployed it, the model provider, or some combination? These questions are still largely unresolved.

Early public-sector moves, such as the UAE’s plan to convert half of federal government operations and services to agentic AI models within two years, underline the urgency of developing robust governance frameworks. When governments themselves rely on agentic systems as “executive partners” in their operations, the need for transparency, auditability, and clear lines of accountability becomes even more pressing.

In the private sector, partnerships like that between TCS and Rezolve AI for agentic commerce, or regional banks’ experimentation with Payman AI for conversational banking, show that mainstream financial institutions are willing to explore agentic workflows, but typically within tightly controlled sandboxes. Many of these deployments keep agents inside walled gardens and rely on human-in-the-loop approvals for critical steps. In crypto, where users often prefer self-custody and open access, finding equivalent controls without undermining openness will be a central design challenge.

Academic and industry efforts around AgentOps and agentic engineering are likely to feed into emerging best practices and, eventually, regulation. For now, the most prudent path combines conservative exposure of agentic systems to real funds, layered defenses around key management and transaction limits, and strong audit trails anchored onchain where possible.

How Crypto Builders Can Engage with Agentic AI

Given the stakes and opportunities, builders across the crypto stack—from base layers and infra projects to exchanges, wallets, and app developers—are actively exploring how to incorporate agentic AI into their roadmaps.

Protocol and infrastructure teams can start by making their systems easy for agents to integrate with. This means offering clear, machine-readable APIs, supporting emerging protocols like MCP and x402 where appropriate, and providing SDKs that encapsulate best practices for secure agent interaction. Chains like Base, Aptos, and Orbs provide examples of how to position infrastructure for agentic workloads, whether by optimizing for low-cost, high-throughput payments, high-performance trading flows, or cross-chain execution verification.

Exchanges and wallets can experiment with agentic copilots and automation features, while keeping strict controls around what agents are allowed to do. Joules’ argument that exchanges must adopt agentic AI or risk failure reflects a broader shift in user expectations: as agents become more capable in other domains, traders and investors will expect similar assistance in managing their crypto assets. However, meeting that expectation requires careful design so that agents cannot bypass security controls, encourage reckless leverage, or violate regulatory requirements.

Developers and founders can look to ecosystems like Animoca’s Minds, the Build East demo day, and Aptos’ ecosystem funding as indicators of where capital and support are flowing. Projects that treat agentic AI as a core product layer rather than a bolt-on feature—and that design their architectures around agents from the outset—are likely to be better positioned as the agentic web matures. Machine learning commentators note that agent cost optimization, interoperability via protocols like MCP and A2A, and robust AgentOps are key differentiators in this emerging landscape.

Outlook

Agentic AI is moving rapidly from concept to reality across science, enterprise software, and financial markets, and crypto is emerging as a natural proving ground for its most ambitious applications. The combination of programmable money, transparent ledgers, and composable protocols provides a uniquely suitable substrate for AI agents that need to reason about value, execute transactions, and be held accountable for their actions.

In the near term, we can expect to see more focused use cases like Travala’s agentic travel protocol, institutional trading infrastructure on chains like Aptos and Orbs, and targeted deployments in payments and banking through platforms such as Payman and x402-enabled rails. These systems will likely keep humans in the loop at key decision points, emphasizing safety and compliance even as agents handle more of the operational burden.

Over the medium to long term, the concept of an agentic web—Web4—may take shape as persistent AI agents with onchain identities and assets become common, interacting across domains from DeFi to gaming, commerce, and public services. If that vision materializes, questions of governance, regulation, and economic design will loom large, and crypto’s experience with decentralized coordination and verifiable state will be invaluable.

For now, the most realistic stance is one of cautious experimentation. Agentic AI in crypto is neither a passing hype cycle nor a fully mature technology; it is an evolving frontier where intelligence, autonomy, and programmable value intersect. Builders who engage early, prioritize safety, and design with both human and machine actors in mind will help shape not just new products, but potentially a new layer of the internet itself.