Binance: AI Trading Agents Under User Control

Binance launches Agent OS, a platform that connects AI agents to user accounts for market analysis and order execution. Permissions are set through sub-accounts and approvals, with withdrawals blocked by default. Operations via the Agentic Wallet are capped each day, while the reasoning of the agents remains out of the exchange's view. Other exchanges, from Kraken to Coinbase, have already initiated similar efforts.

Binance sets daily limits on the Agentic Wallet

Agent OS also connects AI agents to payments and the blockchain: through the x402 integration, they can send and settle payments, and the Agentic Wallet provides access to tokens and decentralized finance protocols. These uses are governed by daily caps set by Binance: $50,000 per day for regular exchanges, a default limit of $100,000 for DeFi, and $20 per day for x402 payments. In contrast, on the trading side, Binance does not impose a separate limit on an agent; the effective cap corresponds to the funds placed by the user in the dedicated sub-account.

Binance's visibility into agents' reasoning is limited

Jeff Li indicates that the reasoning that leads an agent to act occurs outside of Binance's systems, on the user's machine or within the AI application used. He adds that the exchange cannot see this reasoning, even though it can monitor the resulting trading activity. The platform thus acknowledges limited visibility into the influence of potential misinformation or manipulations. In light of the risk of prompt injection or agent compromise, Li presents the sub-account as the first line of defense, and Binance specifies that its security, risk control, and anti-money laundering policies associated with sub-account APIs apply to Agent OS from the launch.

Users configure access through dedicated sub-accounts

Binance places a significant amount of control in the hands of users, who decide on the access and instruments their agents can use, as well as operational limits. Jeff Li emphasizes granular access control at the account level to protect funds. Specifically, sub-accounts can be assigned to agents and configured for specific activities such as spot or futures trading, with withdrawals blocked by default. Users can require approval on a per-order basis or allow autonomous execution once permissions are set.

Agent OS connects AI agents, Binance APIs, and third-party tools

Launched on Thursday, the Agent OS platform integrates autonomous AI into real money management. It connects applications and AI agents to Binance's financial infrastructure and aggregates its existing components: Binance APIs, Agentic Wallet Hub, x402 API, and Skill Hub, with support for the Model Context Protocol. It works with third-party tools like OpenAI's ChatGPT and Codex, Anthropic's Claude Code, or Cursor, allowing access to market data, account information, and order execution. Trading is among the first targeted use cases, and Jeff Li also mentions market monitoring, research, risk analysis, responding to signals, and the autonomous execution of strategies, including arbitrage.

Kraken, Coinbase, and OKX also deploy agent access

Jeff Li describes Agent OS as a first step towards a platform intended to host AI applications operating in both crypto and traditional markets. This trend encompasses the entire sector: in March, Kraken made available an open-source command-line tool with an integrated MCP server, allowing AI agents to perform operations in both the spot market and derivatives. In June, Coinbase unveiled Coinbase for Agents, a solution enabling AI agents to be directly connected to user accounts for trading, payment, and other financial operations, within the restrictions set by clients. OKX, for its part, activated agent trading by introducing an open-source MCP kit earlier this year.