The DSP Is Becoming an Agent: Why Cross Platform Agentic Buying Wins

×

Last week at unBoxed, Amazon retired the name "Amazon DSP."

That sentence would have sounded crazy two years ago. The Amazon DSP was one of the biggest buying platforms in the business. Now its console and the sponsored ads console have been merged into one product called the Amazon Ads Agent (AdExchanger). The traditional DSP setup still exists, but it now sits behind a toggle in advanced settings. The default is an AI campaign type that picks the strategy for you based on your objective.

I don't read this as a rebrand. I read it as Amazon telling the market what a DSP is going to be from here on. Not a seat you log into. An agent you hand an objective to.

And I think it points somewhere even bigger than Amazon.

The DSP is turning into an agent

For fifteen years the DSP business was built around the console. Line items, bid settings, frequency caps, supply paths. The more knobs you exposed, the more serious the platform looked, and the more trader hours it took to run it.

Amazon's announcement flips that. Look at what the Amazon Ads Agent now does in one workflow:

• Generates video creative inside the setup flow
• Builds a media plan automatically, with the advertiser's approval
• Optimizes the campaign once it's live
• Answers performance questions in plain language, so nobody needs SQL to use the data

Amazon says planning and forecasting are next, and they're in development this year. Once that's in place, the agent can go from optimizing one campaign straight into setting up the next one, with the account owner's consent. That's the full loop. Plan, buy, optimize, learn, repeat.

Amazon is also using the agent to fix an old problem. Its properties (Twitch, Fire TV, the Amazon site and app, the audience network) never really fit together as one platform. The agent is the glue. Kelly MacLean, who runs science, engineering and product at Amazon Ads, now calls it "the simplest experience in the industry."

This is the shift. The value of a DSP is moving from the controls it exposes to the decisions it makes on your behalf.

One platform's agent only sees one platform

Here's the catch. Amazon's agent is brilliant at Amazon. It will plan Amazon, buy Amazon, optimize Amazon and attribute back to Amazon. That's its job.

But no advertiser lives inside one platform. A real plan runs across Amazon, Meta, TikTok, CTV, the open web and retail media, all at once. Every one of those platforms is going to ship its own agent. Soon every walled garden will have an assistant that's happy to take your whole budget and grade its own homework.

So the question changes. It's no longer "which DSP do I log into." It's "who decides how the budget moves between all these agents."

That can't be one of the platforms. It has to be a layer above them that has no stake in which one wins. One plan, set up once, translated into each platform's native language, with results coming back into one place where they can actually be compared.

That's cross platform agentic buying. And it's where the DSP business is heading.

The agent is the interface. The infrastructure is the product.

A few weeks ago Aperiam gave this layer a name: the Unified Media Platform, or UMP. I wrote about what it takes to build one, and Amazon's news makes the argument even sharper.

The LLM is not the moat. Everyone is buying from the same handful of labs. If your cross platform product is a chat window sitting on top of a few platform APIs, you've automated the clicking. Useful, but someone else can stand that up in a quarter.

What's hard is the data and model layer underneath the agent. It has to do two things.

1. Turn bespoke campaign setup into platform native execution. Amazon, Meta and TikTok all have different object models, bidding logic and audience rules. Getting from one media plan to each of them is translation, not transcription. Do it badly and you end up with the lowest common denominator campaign everywhere.

2. Give every platform one shared understanding of the campaign. A campaign isn't five campaigns that happen to share a budget. It's one plan expressed contextually, against audiences, and natively inside each platform. If those pieces don't share a definition, the data coming back can't be normalized into a single measurement layer, and you can't tell which agent is actually earning its money.

That layer has to be built from scratch to work across platforms, and it has to stay flexible enough to pull first and third party data into the decision itself, not just the reporting. That takes years, not quarters.

Where Neural Ads started: Amazon, through MCP, on day one

We didn't arrive at this after Amazon's announcement. We built Neural Ads for it from the start, and Amazon was our first MCP integration.

That choice was deliberate. We believed early that buying would move from consoles to agents, and that the connection between a decisioning layer and a platform would be a protocol, not a login. So we built MCP first. Amazon came first because it had the hardest version of the problem: retail signals, a DSP, sponsored ads, CTV and audio, all in one ecosystem that didn't yet talk to itself.

From there we kept going. Today the Neural Ads MCP server exposes more than 100 tools, and the same plan can activate across Amazon, Meta, TikTok, X, Yahoo DSP and Magnite ClearLine. Under the agent sits the part that's actually hard:

• Planning first. Marketing mix models that start from category baselines and refit on the brand's own data, plus reach and frequency forecasting by channel. This is the step Amazon says it's building next. We do it across platforms.
• Blocks as the common unit. Our audiences are built as blocks from programmatic and CTV viewing data together, so one audience definition travels to every platform instead of being rebuilt in each one.
• One decisioning engine. Our Bayesian Incremental Decisioning engine reads results from every platform on the same terms and moves budget toward what's actually incremental.

Amazon's agent decides how to spend on Amazon. Neural Ads decides how much should go to Amazon in the first place, and then speaks Amazon's native language to execute it. Those two things fit together. That's the point.

What this means if you buy media

If you're an agency or a brand, a few things change.

Your traders stop being console operators. The platform agents will handle setup and optimization inside their own walls, and they'll get very good at it. The job moves up a level: setting the objective, deciding the split, and holding every agent to the same measurement.

Your stack gets thinner, not thicker. You don't need another seat. You need one layer that plans once and talks to every platform's agent in its own language.

And your data has to work across platforms, not just inside one. The advertisers who win will be the ones who can compare Amazon's results to Meta's results to CTV's results on the same terms, and move money on that basis.

Amazon retiring the DSP name is the clearest signal yet that the old model is done. The next one is cross platform and agentic. We've been building for it since day one, starting with Amazon.

Sources

• Amazon Ads And The Amazon DSP Get An Agentic Rebrand, AdExchanger, September 29, 2026
• What it takes to build a UMP, Neural Ads