She pointed to early internal testing as an example: In one prototype, Amazon asked an agent to generate a path-to-conversion report showing all ad touchpoints before a purchase. Instead of pulling from an existing API, the agent wrote its own code and ran it through Amazon Marketing Cloud, processing more than three years of data. In other cases, agents defaulted to technically correct but outdated API versions—choices a human developer likely wouldn’t make. MCP tools are designed to prevent those missteps by giving agents explicit instructions for common workflows.
“It’s a lot of heavy lifting to teach the agent how all of this works,” Despins said. “The difference with MCP is you’re going to start from the language in which it operates, and then when you add tools, you take common multi-step workflows and you give an instruction manual to it. Then it’s going to spend its time reasoning on things that matter rather than reasoning—and sometimes getting it wrong—on the basics.”
The open beta follows Amazon’s closed MCP pilot announced last year and mirrors similar efforts across ad tech, including AdCP which was announced last year, as platforms push toward standardizing how AI agents interact with ad systems. It also comes as the ad industry shakes off its old habits–breaking away from rigid, manual ad workflows toward agent-led automation. Meanwhile, Amazon has hinted at leaning further into agentic commerce. In Q3 of last year, CEO Andy Jassy said AI agents could expand how consumers shop online, as the company’s ad business grew 24% year over year to $17.7 billion.
Despins noted that Amazon’s ad partners had been asking for this kind of protocol, though she declined to say how many advertisers participated in the closed beta.
