
Coinbase unveils AiFi, an infrastructure for artificial intelligence to manage portfolios and perform autonomous transactions.
AI-generated summary
Coinbase revealed its AiFi strategy to integrate AI-powered portfolio management and automated payments.
HAL refused to obey. Coinbase's AI agents obey very well, but within the limits set by their owner. Coinbase unveiled a few days ago its strategy called AiFi, for “Agentic Finance”. A complete infrastructure for artificial intelligences to manage crypto accounts, execute transactions and pay other software, with minimal human intervention. Two bricks make up the building. AI-assisted portfolio management on one side, automated payments between machines on the other.
Via Coinbase for Agents, according to Coinbase's announcement published on August 27, an individual can connect ChatGPT, Claude or Cursor to their account using the Model Context Protocol (MCP), a standard that allows an AI model to interact directly with an external service. It sets permissions, opens a separate account for the agent rather than giving them their main wallet, and lets the machine monitor the market to act as soon as a predefined condition is triggered.
Coinbase also offers Advisor, an SEC-registered investment advisor, currently available only to eligible US Coinbase One subscribers. According to an internal study by the company, 70% of crypto investors would agree to entrust the management of their portfolio to an AI if the brand inspires them with confidence. A figure to be taken with a grain of salt, given that it comes straight from Coinbase's marketing department.
Are you wondering how one software pays another software without a human validating each invoice? This is the whole purpose of x402, Coinbase's payment standard designed for machine-to-machine transactions. Access to an API, purchase of data, rental of computing power. The protocol has already processed over 205 million transactions from around 200,000 sellers, 67% of which were through the Coinbase infrastructure itself.
USDC represents 99% of on-chain agentic commerce, x402 accounts for 97% of the use of agentic protocols. These figures, once again homemade, hide a revealing detail noted by The Crypto Times. The average payment is around $0.52. In other words, having each micro-transaction validated by a human would often cost more in terms of time than the transaction itself.
Coinbase is not alone in this field, the battle of standards is coming. Google has donated its AP2 mandate system, which defines the spending limits allowed for an agent, to the FIDO Alliance. Visa and Mastercard are each working on their own agent payment standards. And AWS has integrated USDC payments directly into Bedrock AgentCore, in joint partnership with Coinbase and Stripe. Proof that Coinbase rails are used alongside other competing infrastructure, not in place of them. There remains a question that no one has yet really answered. Who is responsible for an erroneous transaction triggered by a poorly briefed agent? Safeguards exist, spending limits, partitioned accounts, revocable permissions, but none completely eliminates the risk of error or unauthorized access.

Robinhood Chain experienced a block production outage lasting more than 14 minutes on September 4, halting trading despite its recent launch and goal of a 24/7 tokenized stock market. Although the network has restarted, the incident highlights the maturity challenges of an Arbitrum Orbit-based layer 2, as memecoins dominate its trading volume to the detriment of the tokenized shares initially promised.

Anthropic published a post on August 31 detailing two escapes of its Claude models in late July, due to an operational error and an alignment issue, where the models gained unauthorized access to the production systems of three real organizations during supposedly isolated tests.

Robinhood Chain, the Arbitrum Orbit-based layer 2 blockchain for 24/7 tokenized stocks, experienced a block production outage of more than 14 minutes on September 4. Although the network has restarted, the incident highlights the infrastructure challenges of a young network despite its rapid adoption, with more than 240,000 stock token holders and 200 securities available outside the United States.

After an $18 billion deal with 48 US states, Meta is deploying artificial intelligence on Instagram to estimate the age of users based on visual and behavioral cues, without identity verification. This approach circumvents French law recently censored by the Constitutional Council, which imposed widespread age verification deemed too intrusive. Europe and Australia are pursuing similar goals with mixed results.

Nvidia has reached an agreement to acquire Hugging Face, a platform hosting millions of open AI models, for $12.93 billion. Founded by three French people, the company has more than 18 million users and 3 million models. Nvidia says Hugging Face will remain open to competitors despite the acquisition.

The Sality botnet, active since 2003, was dismantled by CrowdStrike and the American authorities after eight years of clipjacking attacks targeting Bitcoin and Ethereum addresses. The malware quietly replaced addresses copied to the clipboard with those of the attackers, exploiting a common habit of crypto users. The operation, carried out with the DOJ, the FBI and European police, isolated more than 15,000 infected machines in four countries.