An ad tech specialist can connect agent-based media buying to an existing demand-side platform (DSP) in three ways: give agents access to the DSP through its APIs or MCP server, let a buyer agent agree deals direct with seller agents and activate them in the DSP through a deal ID, or run agent-to-agent deals with no DSP in the path at all. The three patterns are not rungs on one ladder, and most teams will run more than one at once. The choice depends on where the inventory sits, which protocol the seller's agent uses and how much of each deal the team wants a person to approve.

TL;DR: agents working through a DSP change who operates the platform, agents dealing direct with seller agents change how a deal ID is produced, and agent-to-agent deals with no DSP change where the record of the deal lives. Each pattern can run on AAMP, AdCP or a platform's own APIs, so the protocol question is secondary to the workflow question. In every pattern, agents can take on RFP chasing, IO paperwork and reconciliation while people keep the judgement and approve the deals that matter.

One platform that does not tie a team to a single pattern makes a concrete example. Alkimi is neutral on standards: it can work with agents built on AdCP or AAMP, and it can work with or without a DSP. Its Concourse product benchmarks any buying agent without asking which standard it uses, its Gym lets buyers and sellers build their own agents and spar them against counterparty agents before any real money moves, and its Marketplace gives agents a neutral place to trade, with people keeping approval rights at the deal level.

What are the three ways to connect buying agents, side by side?

The decision table compares the three patterns on the questions an ad tech team has to answer before the first agent goes live.

Question	1. Agents through an existing DSP	2. Agents direct with seller agents, alongside the DSP	3. Agent-to-agent, no DSP in the path
How it connects	The DSP's own APIs or MCP server, or agent services hosted inside the platform as AAMP's ARTF describes	Buyer agent talks to seller agents over MCP, A2A or REST, and the agreed deal returns as a deal ID the DSP activates	Buyer and seller agents agree and book through a shared protocol such as AAMP or AdCP, and the seller's ad server delivers
Where terms are agreed	Inside the DSP, against inventory it already reaches	Between the two agents, before the DSP is involved	Between the two agents, end to end
What stays with people	Strategy, budget limits and sign-off on plan changes	Approval of each deal before booking, and the check that the DSP setup matches it	Approval of commitments above agreed thresholds, and disputes the agents cannot settle
What changes for the ad tech team	Agent permissions, scoped credentials and a log of what each agent changed	Matching agreed terms to deal ID configuration across SSP and DSP	Keeping one shared record of the deal, plus delivery reporting from the seller side
Suits teams when	Most spend is programmatic and already runs through one DSP	The team wants PG, preferred deals or PMPs with named sellers	Inventory sits outside programmatic, such as sponsorships, broadcast or out-of-home

How do agents work through an existing DSP?

In the first pattern, the DSP stays where buying happens and the agent becomes another operator of it. Yahoo DSP launched its Agent Network in June 2026 as "an open framework that connects advertisers with AI agents from leading technology partners directly within Yahoo DSP workflows", letting advertisers bring their own AI, use native agents or combine both "via open MCPs and APIs". Amazon Ads opened a beta of its MCP server in February 2026, which Paula Despins, Amazon Ads' VP of ads measurement, described to AdExchanger as a "translation layer" between existing software and agents.

Standards reach the same pattern from inside the platform. IAB Tech Lab's Agentic Real Time Framework (ARTF), part of its Agentic Advertising Management Protocols (AAMP), is a foundation for agent services "which operate within a host platform and that the orchestrating platform can call directly to accomplish a shared goal". The mechanics are covered in What Is the Agentic Real Time Framework (ARTF)?

For the ad tech team, the work in this pattern is access control rather than new integration: scoped credentials, a list of what each agent may change and a log the team can read afterwards. People keep strategy, budget limits and sign-off on any change above an agreed size.

How do agents deal direct with seller agents alongside the DSP?

In the second pattern, a buyer agent negotiates with seller agents outside the DSP and brings the agreed deal back in as a deal ID. IAB Tech Lab's buyer agent reference implementation, part of AAMP, is built to browse seller media kits, negotiate, book deals through OpenDirect 2.1 and "obtain Deal IDs for DSP activation" in platforms including The Trade Desk, DV360 and Amazon DSP. It talks to seller agents over MCP, A2A or REST.

On the sell side, IAB Tech Lab's seller agent can push a finished deal through the Deal Sync API v1.0, returning the OpenRTB deal ID "that DSPs need for activation". The Deal Sync API specification describes a push from the system where a deal was created to a receiving system, such as SSP to DSP, and its first version assumes negotiation has already happened.

AdCP sits alongside the DSP in the same way. A spokesperson for the AdCP consortium told Digiday in October 2025 that "publishers and platforms can run both OpenRTB and AdCP simultaneously", and AdCP's seller integration guide puts the split plainly: "AdCP handles campaign setup; OpenRTB handles impression-level auctions".

The main job for the ad tech team here is the hand-over: the terms the agents agreed must match the terms configured against the deal ID, and the Deal Sync specification encourages validating deal conditions against OpenRTB requests. Set-up steps are covered in How agentic media buying connects to your existing DSP setup and the deal ID mechanics in How Does an Agent-Negotiated Deal Reach Your DSP?

What does an agent-to-agent deal with no DSP in the path look like?

In the third pattern, the buyer agent and the seller agent agree, book and track the deal between them, and the seller's own ad server delivers it. Within AAMP, Agentic Direct provides the orchestration layer for agent interaction and OpenProposal standardises how publishers define products and packages in RFP responses. AdCP describes itself as working "across every transaction model: direct deals, sponsorships, broadcast, and out-of-home, not only programmatic exchange" in its introduction, and its seller guide says "AdCP sends campaign instructions; your ad server handles delivery".

AdCP 3.2 offers two commercial paths, according to its media buy protocol: a published offer the buyer can select and price directly, and a seller-planned proposal the two sides refine before the buyer accepts it. The protocol's own mapping of agency terms ties the insertion order to the accept_proposal task, which is where IO paperwork goes in this pattern.

What changes most for the team is the record. AdCP's introduction notes that payment, legal agreements, settlement, disputes, credits and makegoods "remain account or platform integrations unless the relevant AdCP contract is explicitly present", so with no DSP report to fall back on, both sides need one agreed account of what was bought. People keep approval of commitments above agreed thresholds and the disputes the agents cannot settle.

Which parts of the job move to agents in each pattern?

Across all three patterns, the work agents take on is the work few media buyers would miss: chasing RFP responses, filling in IO paperwork, reconciling delivery against what was agreed and arguing make-good disputes line by line. What stays with people is the judgement: which sellers to trust, what a brief is really for and whether a deal is worth signing.

Earned autonomy sets that line in five stages: Observe, Recommend, Draft, Human-approved action and Bounded automatic action. A team might let agents reach bounded automatic action inside a DSP it already controls, while holding agent-to-agent deals with a new seller at human-approved action until the record shows the agent can be trusted. The stage is set per pattern and per seller, not once for the whole stack.

Benchmark evidence argues for that caution. In Concourse Bench v1, 55 of 96 runs completed a buy, and 95 of 110 signed contracts promised a viewability floor below 70%.

How should an ad tech team choose between the three?

Four questions settle most of the choice, and the answers often point to different patterns for different parts of one plan.

Where does the inventory sit? Supply the DSP already reaches suits pattern one, named sellers on PG or PMPs suit pattern two, and sponsorships, broadcast and out-of-home suit pattern three.
Which protocol do the seller's agents use? Sellers choose AAMP, AdCP or a platform API, so a buying set-up that works with any of them avoids an early bet.
Does the DSP already expose agent access? An MCP server or agent network makes pattern one quick to start.
Who holds the record? If a deal will never appear in a DSP report, agree before the first booking which record both sides rely on.

Standards are still moving, which argues for light integrations. IAB Tech Lab's public comment period on OpenProposal v3.0, part of the AAMP 3.0 work, runs to 10 November 2026, and AdCP, published by AgenticAdvertising.org since its launch in October 2025, reached version 3.2 in October 2026.

Where does a standard-neutral platform fit across the three?

Alkimi does not ask a team to pick one pattern: teams can benchmark any buying agent in Concourse, spar it against counterparty agents in the Gym before any real money moves, and trade in the Marketplace with or without a DSP in the path. In the Marketplace, the DealSheet is written at agreement, timestamped, held by both parties and retrievable by either side independently, which answers the record question that pattern three raises. Teams that want to test an agent first can start a scoping conversation through the Concourse contact form.

Frequently asked questions

Do you need to replace your DSP to use AI buying agents?

No. Agents can work through the DSP's own APIs or MCP server, or agree deals elsewhere and hand them back as deal IDs the DSP activates.

Can a buying agent buy media without a DSP?

Yes, where the seller's agent can take the booking and the seller's ad server handles delivery. AAMP's buyer and seller agent reference implementations and AdCP's media buy protocol both support booking direct with a seller.

Should an ad tech team build on AAMP or AdCP?

Sellers choose which protocol their agents speak, so a buying set-up that can work with either avoids an early bet. Both standards are developed in the open and both are still changing.

What should stay with people when agents connect to the DSP?

Strategy, budget limits and approval of the deals that matter stay with people. Agents can take on RFP chasing, IO paperwork, reconciliation and make-good disputes inside the limits people set.