Binance launches Agent OS for AI trading

Binance launches Agent OS for AI-driven trading, agents can place orders via sub-accounts and external tools, exchange limits withdrawals while leaving strategy and losses to users

Binance has launched a platform called Agent OS that lets AI agents analyze markets and execute trades on users’ behalf, according to TechCrunch. The crypto exchange, which it says has more than 300 million registered users, is positioning the product as plumbing: a way for developers to connect AI applications to Binance’s trading and payments infrastructure.

Agent OS is built around permissions rather than a house “autopilot.” Users can authorize agents to access market data, view account information, and place orders, and can choose whether an agent must request approval for each order or can trade autonomously once access is granted. Binance’s main containment mechanism is structural: agents operate through dedicated sub-accounts that users fund and configure for specific activities such as spot or futures trading, while withdrawals from those sub-accounts are blocked by default.

The exchange is explicit about what it does not see. Jeff Li, Binance’s vice president of product, told TechCrunch that an agent’s reasoning happens outside Binance’s systems—either on the user’s computer or inside the AI application the user chooses—leaving Binance able to monitor trading activity but with limited visibility into how decisions are made. That matters because the first use case Binance is pushing is trading, where speed and scale are the product: agents can monitor markets, conduct research and risk analysis, react to signals, and execute strategies such as arbitrage.

Binance is also widening the surface area beyond simple order placement. Agent OS integrates existing Binance tools and services, including its APIs, a “Wallet Agentic Hub,” and an x402 transaction verification and payment facilitator interface. It also introduces support for Binance’s own Model Context Protocol, and is designed to work with tools such as OpenAI’s ChatGPT and Codex, Anthropic’s Claude Code, and Cursor, TechCrunch reports.

The risk controls described so far are uneven across functions. Binance does not impose a separate cap on how much an AI trading agent can trade or lose; the sub-account balance is the practical limit. By contrast, Binance-set daily limits apply to its agent-oriented wallet functions: regular swaps are capped per day, DeFi transactions have a higher default cap, and x402 payments are limited to a much smaller daily amount, according to TechCrunch.

In practice, the new platform turns “who clicked buy” into an attribution problem. When an agent’s logic runs off-platform and an exchange only sees the resulting orders, the dispute shifts from market risk to permission design, compromised endpoints, and whether the user understood what was delegated.

Binance says withdrawals from agent sub-accounts are blocked by default, but the trades themselves can still happen at machine speed once access is granted.