
New limits cap trade size and daily volume, with auto-stop on portfolio losses, but do not cancel orders or close positions.
AI-generated summary
Coinbase announced Agent Controls for Agents on Nov. 11, a service connecting users' AI tools to their trading accounts. CEO Brian Armstrong promoted the controls on Nov. 13 as a way to make AI use in financial accounts safe.
Coinbase is rolling out Agent Controls for Agents, its service that connects users' AI tools to their trading accounts.
The Nov. 11 announcement adds limits enforced on Coinbase's servers, while leaving users responsible for the trades their agents make.
On Nov. 13, CEO Brian Armstrong promoted the controls as a way to make AI use in financial accounts safe. A pause on new trades does not mean existing orders have been canceled, positions closed, or losses ended.
In its Agent Controls announcement, Coinbase describes a maximum dollar amount for each trade and a cap on total trading over 24 hours, and trades exceeding the volume limit will not go through.
Auto-stop adds a portfolio percentage-loss trigger, and the agent stops placing trades until the user turns trading back on in the Coinbase app.
Coinbase says enforcement runs on its servers, where the agent cannot bypass the limits or change them itself, while users manage the settings through their authenticated Coinbase account. That separates the rules governing execution from instructions given to an AI model.
Product-category permissions remain forthcoming. Coinbase labels the choice of whether an agent may trade crypto, equities, or derivatives as coming soon.
The announcement does not say auto-stop cancels outstanding orders or closes existing positions, and also gives no loss-calculation baseline or valuation frequency, and does not define how the 24-hour volume window is counted.
Coinbase's agent guide treats order management as a set of distinct operations, while its equity guidance warns that receiving an order ID is not proof of a fill and says individual cancellation results must be checked.
The command reference lists canceling an order and closing a position separately, with availability depending on the product.
The practical implication is that a user cannot treat an agent's stopped status as confirmation that exposure has ended. A pause, a successful cancellation, and a closed position answer different questions. The published auto-stop description supports the first of those, without promising the other two.
Portfolio isolation has a futures exception
Coinbase's setup documentation recommends a dedicated portfolio funded with only what the user is willing to risk, with agent access restricted to it. That makes the user's initial allocation and permission choices part of the protection.
US derivatives, including futures, require the default portfolio. A user considering futures cannot assume the same dedicated spot-portfolio boundary applies across every supported asset class.
Supported AI apps connect through remote MCP using Coinbase OAuth, where users approve portfolios. The local command-line interface instead uses a CDP API key scoped to selected accounts, with Trade and Transfer permissions.
The documented Transfer permission allows internal portfolio moves, and Coinbase also maintains a client allowlist, so compatibility alone doesn't grant access.
The product page says crypto, stocks, and futures availability depends on region and account eligibility. Approval before each trade and unattended scheduling also depend on the agent app.
Coinbase's guide points users to Account Security Connections to revoke access, but deleting a local skill or clearing a chat does not revoke it.
Coinbase also warns that AI agents can fail, trading can lose the entire investment, and users assume responsibility for agent transactions and third-party data use. Users still have to decide what exposure they accept, check what has executed, and manage the resulting positions.
AI outlook — possibilities, not facts
Coinbase will release product-category permissions for crypto, equities, and derivatives trading.
Likely · Within weeks

OpenAI introduced a ChatGPT feature letting AI agents make purchases on users' behalf via Visa, Mastercard, and American Express, starting with Pro subscribers in the US. Tokenized credentials protect card details, and spending-limit automation is planned. Juniper Research projects agentic commerce will reach $1.5 trillion by 2030.

Blockstream, led by cryptographer Adam Back, has launched an institutional digital asset custody solution with French technology group Thales, combining Thales' hardware security modules with Blockstream's Secure-MPC technology to keep signing operations offline and protected from cyberattacks.

Chinese developers are planning and building AI data centers worth over 23 billion yuan ($3.2 billion) around Nvidia processors restricted by U.S. export controls, according to a Reuters investigation. Projects in Fujian, Shaanxi, Jilin, Inner Mongolia, and Wuxi specify H100, H200, or H800 chips, though documents do not confirm all processors were obtained or that export laws were broken.

Amazon released AWS Emulator on Nov. 11, a local tool that imitates AWS services for cloud application testing. The company says autonomous AI agents wrote and reviewed its implementation, while human engineers manage fidelity tests, difference detection, issue fixes, and shipping guardrails. AWS warns the emulator is not for production workloads and does not validate application security.

Forescout researchers used Anthropic's Claude AI to uncover 66 previously unknown vulnerabilities in embedded networking software. The AI-assisted approach reduced analysis time significantly, though researchers noted limitations in detecting complex logic and design flaws.

Google introduced the Gemini Decisions API on Nov. 11, 2026, allowing software to request bounded AI judgments with fixed choices and probabilities, bypassing prose generation.