TL;DR: Agent-to-agent advertising negotiation is live in a small number of environments and in various stages of readiness across a larger set of platforms. This piece maps what is actually available: the protocol requirements for direct negotiation, the integration steps buyers need to take, what each category of platform can and cannot do today, and the limitations buyers should build into their expectations before committing budget.
The straightforward version of agent-to-agent advertising is this: a buy-side agent and a sell-side agent exchange structured messages, agree on terms, and generate a record of the deal without a human intermediating the process. The buy-side agent operates from a mandate approved by the buyer; the sell-side agent operates from an inventory policy approved by the publisher; and the resulting deal record is held in a form that both parties can access and reconcile against.
That is the model. The market reality in 2026 is that some platforms support it, some are building toward it, and some are using "agentic" to describe conventional automated buying with an API layer on top. The distinction matters because the integration requirements, the audit capabilities, and the contractual protections are different depending on which category you are dealing with.
What does a platform need to support genuine agent-to-agent negotiation?
Five capabilities are required. A platform that lacks any of them can support automated buying but not agent-to-agent negotiation in the sense defined above.
The first is a mandate interface. The platform must accept a structured mandate from the buyer: the parameters within which the agent is authorised to act, the approval thresholds that trigger escalation to a human, and the constraints on inventory, pricing, and audience. Without a mandate interface, the agent is operating without defined scope, and there is no way to audit whether its deals were within buyer-approved limits.
The second is a bilateral negotiation channel. The buy-side and sell-side agents must be able to exchange offer and counter-offer messages in a structured format. Most existing DSP/SSP integrations do not have this; they have an auction format where the buyer submits a bid and wins or loses. Bilateral negotiation produces a different kind of deal record because both parties have agreed on terms rather than one party winning a one-sided process.
The third is a shared deal record. The agreed terms must be captured in a record that both parties can access and that neither party can modify after execution. This is technically a state management problem, and it is solved differently by different platforms. Some use a shared ledger; others use an escrow model where the record is held by an independent party; others hold the record at the marketplace layer and provide read access to both parties.
The fourth is a mandate-to-deal audit capability. The buyer must be able to verify, post-execution, that the agent acted within the parameters of the mandate. This requires that the deal record and the mandate record share a reference that allows them to be compared.
The fifth is an approval workflow. Where the agent cannot proceed without human sign-off (because a deal exceeds an approval threshold or falls outside mandate parameters), the platform must be able to pause the negotiation, surface the decision to the appropriate human, receive an approval or rejection, and log the outcome in the deal record.
What categories of platform are available?
The market can be divided into three functional categories. Understanding which category a vendor falls into determines both the integration requirements and what you can do with the resulting deal records.
The first category is agentic marketplaces. These are platforms built specifically for bilateral agent-to-agent negotiation. They have native mandate interfaces, bilateral negotiation channels, and shared deal records. The integration requirement is higher than for conventional programmatic platforms because they require the buyer to provide a structured mandate rather than simply routing a bid. The audit capability is also higher: a correctly implemented agentic marketplace produces a complete mandate-to-deal-to-delivery audit chain.
The second category is agent-ready programmatic platforms. These are established DSPs, SSPs, and ad exchanges that have added API layers allowing buy-side agents to submit structured requests, but whose underlying deal structure remains auction-based. The buy-side agent can automate bid submission, budget pacing, and creative selection, but the "deal" is an auction win rather than a bilaterally negotiated agreement. The audit trail covers delivery but not bilateral agreement, because there was no bilateral agreement in the technical sense.
The third category is workflow automation platforms. These are tools that wrap existing programmatic interfaces with agent-driven automation, handling the human steps in a conventional buying workflow (trafficking, reporting pulls, pacing alerts) without changing the underlying deal structure. These platforms have significant operational value but are not agent-to-agent advertising platforms in the negotiation sense.
Most of the product releases in 2026 that use the word "agentic" fall into the second or third category. That is not a criticism; both have real utility. It is a classification point that buyers need to be clear on before they start an integration.
What are the integration steps for a buyer entering agent-to-agent negotiation?
Integration into a genuine agentic marketplace typically involves four steps.
The first step is mandate definition. The buyer, working with their agency and legal team, must define the parameters under which their buy-side agent is authorised to act. This includes price ranges, inventory constraints, audience parameters, and approval thresholds. This step is not a technical task; it is a governance task, and it requires sign-off from the appropriate budget owner.
The second step is technical integration. The buy-side agent must be connected to the marketplace via API. Most platforms provide a sandbox environment for testing the mandate interface and the negotiation channel before going live. The technical integration includes setting up the approval workflow, which means connecting the platform's escalation triggers to the buyer's internal approval system.
The third step is a pilot run. Before committing production budgets, buyers should run a controlled pilot: defined spend, defined campaign, specific mandate parameters, with the audit chain checked at each step. The pilot should confirm that the mandate-to-deal reconciliation works correctly, that the approval workflow triggers and resolves as expected, and that the deal record can be exported in a usable format.
The fourth step is ongoing mandate governance. The mandate is not a set-and-forget document. As the campaign evolves and the agent's performance is evaluated, mandate parameters should be reviewed and updated through the same governance process that approved them initially. An agent operating on an outdated mandate is operating on stale authorisation.
What are the current limitations?
Four limitations are worth building into buyer expectations now.
Inventory access is constrained. The publishers and SSPs connected to agentic marketplaces are a subset of the addressable programmatic universe. Buyers who rely on specific publisher relationships or who need access to specific inventory sources may find that those sources are not yet connected to a bilateral negotiation environment.
Protocol fragmentation is real. The two competing specifications (AdCP and AAMP) create integration overhead for any buyer that wants to work across multiple agentic platforms. An agent built for one protocol does not automatically work with a platform using the other. This will resolve as the market converges, but convergence has not happened yet.
Mandate tooling is immature. Most platforms have a mandate interface, but the tooling for managing, versioning, and auditing mandates over time is underdeveloped. Buyers running multiple campaigns with different parameters across different mandates are managing more complexity than the current tooling makes easy.
Human oversight at scale is an unsolved problem. As agent-negotiated deal volume grows, the number of escalation events requiring human approval grows with it. The approval workflow infrastructure most platforms provide works well at low volume. At high volume, the governance model needs to be designed carefully to avoid either bottlenecking the agent on human approvals or granting approvals too quickly for genuine human review.
None of these limitations are permanent. They are characteristics of a market at an early stage of infrastructure maturity. Buyers who understand them are better placed to build a programme that works within them now and scales as they are resolved.