Coinbase gives AI agents their own crypto wallets

Today’s AI can write code, draft emails, and analyze dense documents, but the moment a real-world payment is required, it hits a wall. Machines simply do not have bank accounts.

Coinbase wants to change that by handing software agents the keys to crypto.

On Wednesday, Coinbase developers Erik Reppel and Josh Nickerson unveiled Agentic Wallets, an infrastructure suite purpose-built for autonomous AI agents. The setup allows AI to independently earn, spend, and execute trades onchain on behalf of users.

Giving machines an autonomous balance sheet

The new tool builds directly on Coinbase’s AgentKit framework launched back in November 2024. Operating on the Ethereum Layer-2 network Base, these agents are designed to run around the clock.

Under the hood, the transactions run through x402, a specialized payment protocol Coinbase built for autonomous machine use cases. According to the company, x402 has already logged 50 million transactions.

Using this pipeline, agents can independently buy API keys, rent compute, tap into premium real-time data feeds, and pay for decentralized storage. They can also manage DeFi liquidity positions without human intervention.

If an agent spots a more profitable yield at 3 a.m., it can rebalance the position instantly, provided the user has configured the trading boundaries and spending allowances beforehand.

Coinbase is far from alone in this race. On the same day, Lightning Labs rolled out a toolkit enabling AI agents to transact on Bitcoin’s Lightning Network via the L402 protocol, managing native BTC without ever holding the private keys. Earlier in the week, Crypto.com chief Kris Marszalek launched ai.com for task-oriented personal agents, while Google recently rolled out its Universal Commerce Protocol using Google Pay for dollar transfers.

Why it matters

We think autonomous AI agents might just be the most legitimate use case crypto has seen in years.

The legacy banking system simply cannot handle software agents. Banks are built for humans with government IDs, and processing sub-cent machine-to-machine micropayments breaks the economics of traditional payment processors. If millions of AI agents need to pay each other for compute, data, or digital services, permissionless blockchains and stablecoins are practically the only viable infrastructure.

Circle chief Jeremy Allaire recently predicted billions of agents will use stablecoins for payments within five years, while former Binance head Changpeng Zhao argued crypto will naturally serve as the default currency for AI.

For developers and power users, this turns passive software into active economic participants. Instead of writing rigid, brittle automation scripts, you can direct an LLM to optimize yield or keep server infrastructure funded on its own.

Still, giving an autonomous model direct access to capital is inherently risky. Between LLM hallucinations and prompt injection vulnerabilities, a compromised agent could drain funds in seconds via irreversible blockchain transactions. Setting strict spending caps and whitelisting trusted addresses will be non-negotiable for anyone experimenting with this today.