In the span of two weeks, three major crypto infrastructure companies — Coinbase, MoonPay, and Phantom — each shipped dedicated wallet infrastructure for AI agents. This is not a coincidence. It is the opening salvo of a platform war for the most consequential new user class in crypto: autonomous...
"What happens if there's not a human behind it at all? It's some piece of code that owns a wallet, executing code to make more money… How does liability work in that case? I actually don't know." — Avichal Garg, Partner, Electric Capital
In the span of two weeks, three major crypto infrastructure companies — Coinbase, MoonPay, and Phantom — each shipped dedicated wallet infrastructure for AI agents. This is not a coincidence. It is the opening salvo of a platform war for the most consequential new user class in crypto: autonomous software.
Coinbase launched Agentic Wallets on February 11. MoonPay followed with MoonPay Agents on February 24. Phantom debuted its MCP Server on February 18, enabling AI agents to sign transactions and swap tokens across chains. Meanwhile, Ethereum's EIP-7702 — now live on mainnet via the Pectra upgrade — provides the permission framework that makes scoped, temporary agent delegation possible at the protocol layer. The infrastructure stack for machines that spend money is being assembled in real time.
But the economic reality check is sobering. Coinbase's x402 payment protocol — the open standard designed to be HTTP's native payment layer for machine-to-machine transactions — saw daily activity collapse from 731,000 transactions per day in December 2025 to roughly 57,000 by February 2026, a 92% decline. The plumbing is being built. The demand has not yet arrived. This gap between infrastructure investment and actual usage defines the current moment — and will determine whether the "agent economy" becomes crypto's next genuine value layer or its next narrative bubble.
The coordinated timing of launches across Coinbase, MoonPay, and Phantom in February 2026 reveals a shared bet: that AI agents will become a primary consumer of blockchain infrastructure within 12–18 months.
Coinbase Agentic Wallets (Feb. 11): Purpose-built wallet infrastructure enabling AI agents to autonomously spend, earn, and trade crypto. All private keys are routed through trusted execution environments — agents never see or store keys directly. Session keys enforce spending limits and scope what an agent can do, containing damage if a model behaves unpredictably. The wallets ship with pre-built action libraries — Authenticate, Fund, Send, Trade, Earn — eliminating the need for developers to build transaction logic from scratch. Built on Base, Coinbase's Layer 2.
Phantom MCP Server (Feb. 18): Phantom, the dominant Solana wallet with significant multi-chain reach, launched a Model Context Protocol server that allows AI agents to quote swaps, sign transactions, and transfer tokens across all Phantom-supported networks. Compatible with Claude, OpenClaw, and other MCP clients, the server supports per-action prompts, parameter bounds, and post-execution receipts. Users must grant explicit permissions and can revoke at any time.
MoonPay Agents (Feb. 24): A non-custodial software layer enabling AI agents to create wallets, manage funds, and trade on behalf of verified users. Once a user completes identity checks and funds a wallet, an AI agent can trade, swap, and transfer assets independently. The system includes recurring purchases, real-time cross-chain swaps, machine-to-machine payments, and automated fiat on-ramps. MoonPay CEO Ivan Soto-Wright framed it directly: "AI agents can reason, but they cannot act economically without capital infrastructure."
These are not experimental toys. These are production-grade financial rails, built by companies that collectively process billions of dollars in crypto transactions annually, all aimed at the same thesis: software will soon need its own bank account.
Underpinning much of this infrastructure is x402, the open payment protocol Coinbase co-developed to repurpose the long-dormant HTTP 402 "Payment Required" status code for instant blockchain settlements. The concept is elegant: when a client makes an HTTP request to a resource server, the server can respond with a 402 status code and payment requirements. The client's agent pays — in stablecoins, on-chain — and the transaction completes. No API keys. No subscription management. No human approval.
The protocol launched in May 2025 alongside collaborators including AWS, Anthropic, Circle, and NEAR. By September 2025, Coinbase and Cloudflare co-launched the x402 Foundation to promote adoption. Partners including Stripe and Cloudflare are utilizing it for pay-per-use API models. By December 2025, x402 had processed over 75 million cumulative transactions worth $24 million — primarily for paid APIs and AI agent services.
This is the protocol layer that converts the internet itself into a marketplace where machines pay machines. And it represents a potentially massive economic value layer: every API call, every compute resource, every data feed could become a micro-transaction settled on-chain in stablecoins.
The gap between infrastructure ambition and current demand is stark.
x402 daily transactions cratered from approximately 731,000 per day in December 2025 to roughly 57,000 by February 2026 — a decline exceeding 92%. BeInCrypto's analysis identified a sharp contraction in the "Infrastructure & Utilities" segment as the primary cause, with activity from bridge services like x402secure.com, agentlisa.ai, and pay.codenut.ai dropping over 80% from highs.
The AI agent token market tells a similar story. CoinGecko lists over 550 AI agent crypto projects with a combined market cap of approximately $4.34 billion — down from over $40 billion in early 2025. The token speculation ran far ahead of the infrastructure; now the infrastructure is arriving, but the speculative capital has retreated.
This pattern is structurally familiar to anyone who has followed crypto cycles: the narrative bubble inflates, collapses, and then the actual builders ship during the quiet period. The question is whether this infrastructure phase will produce self-sustaining economic activity — real revenue from real machine-to-machine payments — or whether it will remain, like so much of blockchain, a subsidized experiment.
Through the economic value lens, the critical metric is not transaction count or token market cap. It is fee revenue. How much are AI agents actually paying to use blockchain rails? How much of that fee revenue accrues to infrastructure operators versus being subsidized by token incentives and VC capital? Until agent-driven transactions generate meaningful, recurring on-chain fee revenue that exceeds the cost of maintaining the infrastructure, the agent economy remains in the subsidy phase — no different from the 85–90% subsidy dependence that characterizes the broader blockchain ecosystem.
The legal implications are as significant as the technical ones. At NEARCON 2026 on February 24, Electric Capital partner Avichal Garg drew a striking historical parallel: the creation of the limited liability corporation in the 19th century — a legal innovation that unlocked pooled capital and industrial-scale growth. AI agents with wallets, he argued, represent a similarly foundational shift.
But the comparison also exposes the problem. LLCs work because they have legal personhood, defined liability structures, and regulatory frameworks. AI agents have none of these. "You can't punish an AI," Garg noted. "You can turn them off, but they don't care."
Fenwick & West, a major Silicon Valley law firm, has flagged several unresolved regulatory questions: if an AI agent autonomously executes a trade that constitutes market manipulation, who is liable — the developer, the deployer, the operator, or the company that benefits? If an agent manages assets on behalf of users, does it trigger money transmitter licensing requirements? If it executes trades autonomously, is it operating as an unregistered broker-dealer?
The proposed solution stack — spending limits, policy-based execution, audit logs, attribution systems — is pragmatic but incomplete. It mirrors the "guardrails" approach that Coinbase has built into Agentic Wallets: session keys, spending caps, trusted execution environments. These are risk containment mechanisms, not legal frameworks. The regulatory apparatus has not yet caught up with the technology, and the gap between what agents can technically do and what the law permits is widening with each product launch.
The strategic logic behind the February blitz becomes clear when you follow the value capture dynamics.
Coinbase is channeling agent activity onto Base, its proprietary L2. Every agent transaction that settles on Base generates sequencer revenue for Coinbase. Agentic Wallets are the on-ramp; Base is the tollbooth. This is the Coinbase flywheel: developer tools attract builders, builders attract agents, agents generate transaction volume, transaction volume generates fees.
MoonPay is positioning as the fiat on-ramp for the agent economy. Its competitive moat is regulatory compliance infrastructure — KYC/AML — which is precisely the bottleneck that autonomous agents cannot solve on their own. If every agent needs a verified human behind it (as current regulations require), MoonPay's identity layer becomes mandatory infrastructure.
Phantom is betting on the multi-chain agent. Its MCP Server works across all Phantom-supported chains, with particular strength on Solana — the ecosystem that already hosts the densest DePIN and AI-adjacent infrastructure. By becoming the signing layer for AI agents, Phantom positions itself to capture the cross-chain agent routing market.
Each company is fighting for a different chokepoint in the same value chain. The winner will be determined by which chokepoint proves most critical as actual agent demand materializes.
Three major infrastructure launches in 14 days — Coinbase (Feb. 11), Phantom (Feb. 18), and MoonPay (Feb. 24) — signal a coordinated industry bet on AI agents as crypto's next primary user class.
x402 protocol activity collapsed 92% from December 2025 to February 2026, revealing a stark gap between infrastructure supply and agent demand. The plumbing exists; the customers do not — yet.
The legal liability framework is missing. AI agents can now autonomously hold assets, execute trades, and pay for services, but no jurisdiction has established who bears responsibility when autonomous code causes financial harm.
The platform war is about chokepoints: Coinbase wants agents on Base (sequencer revenue), MoonPay wants to own the fiat-to-agent identity bridge (compliance moat), and Phantom wants to be the multi-chain signing layer (routing fees).
The economic value test remains unmet. Until agent-driven transactions produce self-sustaining fee revenue exceeding infrastructure costs, the "agent economy" is another subsidized narrative — indistinguishable from the token-incentive cycles that characterize 85–90% of blockchain's current economic activity.
Industry forecasts project a $30 trillion autonomous agent economy by 2030. The distance between that projection and today's 57,000 daily x402 transactions is the measure of both the opportunity and the risk.
The agent wallet infrastructure race of February 2026 is a bet on a future that does not yet exist in economic terms. The technology works. The protocols are live. The wallets are shipping. But the demand — the actual, revenue-generating, fee-paying machine-to-machine activity that would validate the thesis — remains largely theoretical.
This does not mean the bet is wrong. The convergence of increasingly capable AI models, mature blockchain payment rails, and standardized protocols like x402 creates a plausible path to an economy where autonomous software is a meaningful on-chain participant. Forty cents of every crypto venture capital dollar in 2025 went to AI-related companies, up from eighteen cents in 2024. The capital is flowing.
But the lesson of every previous crypto infrastructure cycle applies: building the highway does not guarantee the traffic. The agent economy will prove itself not through product launches, token market caps, or venture funding rounds, but through one metric that cannot be faked — sustained, self-funding on-chain fee revenue generated by machines paying machines for real economic value. That is the test. It has not been passed yet.