On March 4, 2026, Amazon added an Agent Policy to the Amazon Services Business Solutions Agreement, the contract every seller accepts to sell in Amazon's stores. Amazon announced the change in its own seller forum and set three duties for every AI agent: identify itself as an automated system, comply with the Agent Policy at all times, and stop accessing Amazon when Amazon asks.
The same announcement says Amazon "may restrict their access in certain instances under the BSA and new Agent Policy." For an agency whose agents change budgets and bids on client accounts, a restriction arrives in the middle of running campaigns. The agent loses access, and every budget and bid it set last keeps spending until a person changes it.
What Amazon's announcement says
The forum post is short. The parts that concern agents read:
Effective March 4, 2026, we're updating the Amazon Services Business Solutions Agreement (BSA) and including a new Agent Policy with new requirements for AI usage and automated systems.
All AI agents must do the following: Clearly identify themselves as automated systems; Comply with the new Agent Policy at all times; Cease access if Amazon requests.
The Agent Policy itself is a help page in Seller Central, and it opens only after a seller login. We have not read it. This post works from the public forum announcement only, and a seller who can open the policy page should read it before relying on anything here.
Several third-party blogs describe thresholds above which a person has to approve an agent's action, for example price changes of more than 20 % within 24 hours. These figures circulate in secondary sources. We could not confirm them against Amazon's text, and we do not treat them as Amazon's rules.
Where ad agents stand
The BSA governs selling. Advertising runs through a separate interface, the Amazon Ads API, and since February 2, 2026 also through the Amazon Ads MCP Server. MCP, the Model Context Protocol, is a common format through which an AI model calls outside functions as named tools. Amazon announced the server in open beta, "available globally in open beta to Amazon Ads partners with active API credentials". Its documentation lists 110 tools in 18 groups, including tools that change daily budgets, change bids, delete campaigns and edit which users may access an ad account.
The forum post does not say whether the Agent Policy applies to agents that work through the Ads API or the MCP Server. It speaks of AI agents in the context of the BSA and says nothing about advertising. Without the policy text we cannot answer the question either way, and an agency should not assume an answer in either direction.
Many agencies run both sides for a client: listings and prices through Seller Central, campaigns through the Ads API. One agent may then fall under the policy while the coverage of the next one is unclear. Applying the three duties to every agent that touches a client account costs little, so that is how we would plan until Amazon says more.
The three duties, applied to an ad agent
Identify as automated
A call to the MCP Server carries an Amazon-Ads-ClientId header and an Authorization: Bearer token from a Login with Amazon app; OAuth 2.1 has been supported since July 6, 2026. The documentation we read shows no field that marks a call as coming from an agent. An agency can still make its agents traceable on its own side by giving them a Login with Amazon app that no person uses, so that every call made under that app is known to be automated.
Comply at all times
"At all times" includes 3 a.m. on a Sunday, when no account manager is watching. A duty that has to hold without a person present has to be checked in code. In practice this means a written limit per ad account and a check of each tool call against it before the call reaches Amazon. A limit that exists only in the agent's instructions is read by the model and checked by nobody.
Cease access on request
Stopping one agent is easy when it runs alone. Stopping all agents on one client account within minutes, while the same agents keep working on the agency's other accounts, needs a switch per account. Revoking the Login with Amazon token stops everything at once, reporting included, and it has to be undone for every account that was not affected.
What the announcement leaves open
- Whether agents that act through the Amazon Ads API or the Ads MCP Server fall under the BSA Agent Policy.
- What "identify themselves as automated systems" means on the wire, as a header, an app setting or a registration.
- How Amazon sends a request to cease access, and whether it goes to the seller, the agency or the developer of the app.
- Whether the approval thresholds quoted on third-party blogs are in the policy text.
An agency outside Seller Central cannot settle these questions. It can prepare for each likely answer now, because the work is the same in every case: a limit per account, a check per call, and a stop per account.
A written limit per ad account
MolTrust Ads checks the tool calls of agents on the Amazon Ads MCP Server. The agency writes a mandate per ad account: which tools the agent may call, which brands, the daily cap, the account-wide cap, the largest bid change, the amount above which a person has to approve, and the date the mandate ends. Before each call, a wrapper in the agency's own code asks MolTrust for a verdict: allow, hold for approval, or forbid. On allow, the agency's own MCP client sends the call to Amazon.
MolTrust holds no Amazon credentials, sees no Amazon data and never calls Amazon. Tools that change who can access an account (user permissions, invitations, manager account links) cannot be granted in a mandate at all. To stop every agent on one account, the agency sets the mandate's end time to now, and the next call is forbidden. Each verdict comes with a record that anyone can recompute offline with the open-source verifier moltrust-verify.
The check keeps no state of its own. Current budgets, bids and the decision time come from the agency's side, and a call sent to Amazon without the wrapper is not covered. Free covers one ad account and 1,000 checks a month; Pro is USD 49 per ad account per month and can be cancelled in any month.
Next step
Run the Agent Liability Check for one client account. It estimates what one wrong budget costs and asks about caps, approvals and the stop. Every field is optional, and nothing leaves your browser. Engineers can go straight to the quickstart or clone the example repository, which returns a first verdict without touching Amazon.
MolTrust is an independent service provider and is not affiliated with or endorsed by Amazon. Amazon Ads is a trademark of Amazon.com, Inc. or its affiliates.