Article

Designing Hard Limits for AI Trading Agents

Published November 16, 2026 5 min read By the Lotvs team 2 sources, read November 16, 2026

The integration of artificial intelligence into financial services is moving rapidly from passive research to active execution. On November 13, 2026, The Paypers reported that Coinbase introduced Agent Controls for its Model Context Protocol offering. This update allows users to set strict limits on how AI agents trade on their behalf. The release brings research, strategy, and trade execution into a single workflow within the AI tools users already rely on. This shortens the path from an initial idea to a placed order.

However, removing friction from financial transactions requires a corresponding increase in safety measures. Coinbase framed this release as a set of "safeguards intended to protect users and their assets" as trading increasingly takes place through automated tools. The introduction of these controls represents a clear shift in how the industry views agentic commerce. Autonomy without oversight is a liability. By defining the exact boundaries within which an agent operates, platforms can offer the benefits of automated execution while keeping the human user in complete control.

Defining the Boundaries of Autonomy

The new release includes three active controls and one planned addition. Each control targets a specific risk associated with automated trading. The first is a maximum trade size. Users can set a cap in US dollars for each individual order. If a user sets a limit of one hundred dollars, the agent cannot execute a single trade above that amount. This prevents a malfunctioning agent from placing a massive order based on a flawed market analysis.

The second control is a 24 hour volume limit. This restricts the total amount of capital an agent can trade within a single day. Any trade that pushes the total volume past this ceiling simply fails to execute. Volume limits protect users from high frequency trading errors, where an agent might rapidly buy and sell an asset, draining the account through fees or poor timing.

The third control introduces an automatic stop mechanism based on portfolio performance. Users set a percentage loss threshold. If the portfolio value drops below that specific level, the platform halts all agent trading. For example, with a five percent threshold, trading stops immediately once the portfolio is down by more than five percent. The agent cannot resume activity until the user manually turns trading back on within the application.

The fourth control is marked as coming soon. It will introduce product permissions. Once available, this feature will allow users to choose which product categories an agent can trade. The options will include crypto, equities, and derivatives. A user could enable equities while leaving derivatives disabled. This planned feature indicates that the agent offering is set to span multiple asset classes, requiring even more granular oversight.

Separating Intelligence from Authority

The most important aspect of this release is a structural design choice. The controls exist entirely outside the agent. The platform enforces the rules. The agent cannot change its own permissions. It cannot bypass the limits.

This separation addresses a fundamental problem in agentic trading. If a system has the power to place orders, it should never have the power to modify the rules that govern those orders. By keeping the controls at the platform level, the account holder remains the only party capable of setting or adjusting the parameters. Users manage these settings through their logged in account, which is protected by a password and two factor authentication.

Interestingly, the system maintains transparency between the agent and the human user. Users can ask their AI assistant to show them where the controls are located. The agent understands its own boundaries and can guide the user to the correct settings page. This creates a collaborative environment. The agent handles the research and the strategy. The platform enforces the safety checks. The human user holds the ultimate authority.

Agent Trade Approval Flow. A flowchart showing an AI agent requesting a trade, the platform checking limits, and the trade either executing or failing.

What This Means for Builders

Builders creating agentic workflows must study this architecture carefully. The primary lesson is that you must build for failure. Coinbase explicitly states that AI agents may make mistakes or fail. They also clarify that users direct all agent activity and assume the risk for the transactions their agents execute. Users also assume the risk for any use of their data by third party AI providers.

The controls do not shift responsibility away from the user. Instead, they provide the exact mechanisms needed to constrain what an agent is able to do. When you build financial tools for AI, you must separate the decision engine from the permission layer. Do not rely on system prompts to enforce trading limits. Large language models can be persuaded to break their own rules. Hard limits must exist at the infrastructure level, completely inaccessible to the agent itself.

Furthermore, builders must design for human intervention. The automatic stop feature is a perfect example of this principle. When the portfolio hits a loss threshold, the system requires a human to acknowledge the situation and manually restart the process. This friction is intentional. It forces the user to review the agent strategy before more capital is deployed. You should incorporate similar circuit breakers into any product that moves money.

If you are building an API for agents, consider how your endpoints communicate failure. When a trade is rejected due to a volume limit, the API must return a clear error message. The agent needs to understand why the trade failed so it can inform the user, rather than endlessly retrying the same doomed request.

The Future of Human Oversight

The introduction of product permissions shows where the market is heading. Agents will soon navigate complex environments containing multiple asset classes. As the scope of agentic commerce expands, the need for robust oversight will only grow. The combination of per trade caps, volume ceilings, and loss based stops gives users several layers of exposure management.

We are moving toward a model where agents act as highly capable researchers and strategists. They will monitor markets, identify price discrepancies, and prepare orders faster than any human could. But the final execution will always depend on human defined boundaries. The user sets the rules of engagement. The platform enforces them.

This framework builds trust. People will only delegate financial tasks if they know the maximum possible downside. By providing clear volume ceilings and loss based stops, platforms give users the confidence to experiment with AI tools. They can let the agent explore the market, knowing their core assets remain protected by unbreakable rules.

The era of unchecked automation is ending before it truly begins. Financial platforms are recognizing that intelligence requires boundaries. By implementing strict agent controls, the industry is creating a sustainable path forward for automated commerce. Agents will continue to find better prices and execute complex strategies. However, humans will always retain the final say.

Share this article

Next step

Try it with Lotvs

One account for you, and one for each AI agent you connect, with limits you set.

Sources

Every fact above links to one of these. Each was read on November 16, 2026; the date beside it is the page's own.

  1. The Paypers, thepaypers.com, November 13, 2026
  2. CryptoRank, cryptorank.io, November 7, 2026

More from the blog

Guides

  • Converting with your AIAsk Claude or ChatGPT to read your balances, check the rate between dollars and euros, explain what it finds and prepare a conversion. Nothing converts until you approve it in Lotvs.
  • Approvals for AI agentsIn Lotvs, an AI agent cannot spend on its own. Every payment or conversion it asks for waits until you approve it in Lotvs and sign it from your own account. Requests expire, and you can pause or disconnect an agent at any time.

Important

Articles describe products and protocols as their sources did on the dates shown, and those change; check the sources before you rely on them. An AI agent connected to Lotvs has no money of its own. It can read your account, tell you what it finds and ask you to pay someone or to make a conversion. Nothing moves until you approve it in Lotvs and sign it from your own account.

Lotvs is a financial technology company, not a bank. Balances are held as digital dollars, not bank deposits, and are not covered by deposit insurance. Names of other companies and their products belong to them, and Lotvs is not affiliated with them. See the Terms and Privacy Policy.