Limits are enforced by Coinbase rather than the agent
A user manages the settings inside the Coinbase application or website. The agent can read the permitted boundaries, but it cannot raise a limit or turn a disabled product back on. This separation is important because the same model that proposes a trade should not be able to remove the rule that restricts it.
The first control sets a maximum value for each trade. A second control caps the total amount traded during a rolling 24-hour period. Any instruction above those thresholds is rejected by Coinbase even if the agent believes the order is appropriate.
Auto-stop reacts to portfolio losses
Coinbase announced the controls on 11 November. The auto-stop option allows a user to set a percentage loss threshold. If the portfolio falls below that level, the agent cannot place more trades until the user manually reactivates trading in the Coinbase application.
The feature does not guarantee that losses will stop exactly at the selected percentage. Market prices can move quickly, and orders already in progress may execute at different levels. It is a permission boundary, not insurance against volatility or an automatic investment strategy.
Product permissions are planned as a separate layer
Coinbase says users will also be able to choose which product categories an agent may trade, including crypto, equities and derivatives. That option is marked as coming soon. Keeping categories separate can prevent a general instruction from reaching a product with a different level of risk.
The controls apply to Coinbase for Agents, the company’s connection between AI tools and its research and trading workflow. Users remain responsible for instructions and transactions. Coinbase also warns that AI agents can make mistakes or fail, even when limits are active.
The design reflects a wider problem in agentic finance
An AI assistant can move quickly from research to execution, which saves time but also reduces the pause in which a person might notice a bad assumption. Server-side limits add that missing boundary. They can contain the size of an error without requiring the model to recognise its own mistake.
Useful protection will still depend on conservative settings, account security and human review. Passwords and two-factor authentication protect the account, while Agent Controls restrict what an authorised agent can do after connection. Together they create layers of defence, but none replaces careful oversight or changes the risk of trading.
This article is for information only and is not investment advice.
admin_wcContributor · World-CryptMore from
