Agent-to-agent protocols change the shape of marketing stacks
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
Most agentic marketing work so far has been about one agent using software: reading a CRM, drafting into a document, querying analytics. A second category of open standards addresses a different problem, which is how one agent works with another agent built by a different vendor or run by a different company.
These agent-to-agent protocols describe how an agent can advertise what it is able to do, accept a task from another agent, report progress, and return a result. For marketing teams, the important change is less about any single standard and more about architecture. Your stack may soon include agents that coordinate with agents inside your ad platforms, your partners' systems, and eventually your customers' purchasing tools.
In this post we look at the patterns we expect, the questions they raise for marketing operations, and the preparation that makes sense before you adopt any specific protocol.
How agent-to-agent differs from tool use
A tool protocol connects an agent to a capability. The agent decides what to do, calls a function such as "search contacts", and gets back a structured result. Control stays with the calling agent from start to finish. We cover that side in MCP for marketers.
An agent-to-agent exchange hands over part of the decision. When your orchestrator asks a vendor's agent to "build a lookalike audience for this campaign", the vendor's agent chooses its own steps. It may call its own tools, ask clarifying questions, or run for minutes before it answers. Your agent sees the request and the result, but usually not the reasoning in between.
That difference matters for governance. With tools, you control the action surface by choosing which functions to expose. With delegation, you control it by choosing whom to trust, what to ask for, and what evidence you require back.
Three patterns to expect
We expect agent-to-agent communication to show up in marketing stacks in three broad forms, each with a different trust profile.
Internal coordination. Your own agents, built on different platforms, exchange work. A planning agent in one tool asks a content agent in another to produce assets for a launch. A shared protocol replaces custom glue code, and both sides answer to the same organization and policies.
Vendor agents. Software vendors increasingly ship their own agents. Rather than calling a vendor's tools directly, your orchestrator may delegate a task to the vendor's agent, which knows its own platform best. The vendor's agent acts inside your account, so its permissions are your permissions.
External agents. Agents acting for other organizations contact yours. A buyer's procurement assistant gathers product and security information, or a partner's agent coordinates dates and assets for a co-marketing campaign. Here the other side answers to a different company with different goals.
External agents are the hardest case to govern. Preparing for buyer-side agents is the subject of agentic commerce.
New questions for marketing operations
Agent-to-agent communication raises questions that most marketing teams have never had to answer in writing. We find it helps to work through four of them before any pilot:
- Identity: how do you know which agent you are talking to, and which person or company it acts for?
- Authority: what is an external agent allowed to request, and what is your agent allowed to promise on your behalf?
- Data sharing: what information may go to another organization's agent, and under what terms?
- Accountability: when a task passes through three agents and goes wrong, whose log tells the story?
Identity and authority are closely linked. A protocol can tell you that a request came from a particular agent endpoint. It cannot tell you whether that agent's operator is allowed to see your pricing roadmap. That judgment still belongs to your policies, and to the people who set them.
A worked example: a co-marketing request
Consider a hypothetical partner whose agent contacts your campaign agent to plan a joint webinar. It asks for three things: a list of your target accounts in a shared vertical, draft copy for the invitation, and a commitment to split promotion costs evenly.
Each request lands in a different place on your policy. Draft copy is low risk and easy to review, so your agent can prepare it and route it to the partner owner for approval. The target account list is restricted data, and sharing it with another company may raise contractual and privacy questions, so your agent should decline and explain that a person will follow up. The cost split is a commitment, so your agent can summarize the request and the options, then stop.
Notice that none of these responses depends on the protocol. They depend on a written classification of data and a clear line around commitments. The protocol only carries the messages. The same pattern of explicit inputs, outputs, and stop conditions appears in designing agent handoffs, and it applies across company boundaries just as it does inside one team.
Practical steps now
You do not need to adopt any protocol today to prepare for this shift. The following steps pay off regardless of which standards your vendors support:
- Classify what agents may share. A simple scheme such as public, internal, and restricted will be needed under any protocol, and it gives agents a rule they can apply.
- Require logs at every boundary. Each time work crosses from one agent to another, record the request, the response, and the identity of both sides.
- Keep commitments human. Agents can gather information and prepare proposals, but commitments on price, terms, or contracts should require a person.
- Bound delegated tasks. Give each delegated task a scope, a time limit, and a definition of done, so a vendor agent cannot drift into adjacent work.
- Ask vendors direct questions. Which agents does the product run, what can they do in your account, and how are their actions logged and exposed to you?
The Agentic Hub architecture shows where these boundaries sit in a typical system. In most designs, the hub is the natural place to enforce the data classification and record boundary logs, because every delegated task passes through it.
What to measure once agents start talking
Once any agent-to-agent traffic exists, a few simple measures tell you whether your boundaries are holding. Track how many delegated tasks complete within scope, how many requests your agents decline under policy, and how many escalate to a person.
A rising decline rate is not necessarily bad news. It may mean partners are asking for more than your policy allows, which is a conversation for people to have. A decline rate of zero across many requests deserves a closer look, since it may mean your agents are sharing more than you intended.
Also review a sample of boundary logs each week while volumes are low. You are checking whether the recorded identity, request, and response are enough to reconstruct what happened without asking anyone. If they are not, fix the logging before volume grows.
What we are studying next
The open question we keep returning to is trust that travels. Inside one company, you can grant an agent permissions and revoke them. Across companies, each side sees only its own half of the exchange. We are studying how much of the other side's run history two organizations should share after a failure, and what a fair default looks like before anything goes wrong.
How to apply this: write a one-page policy covering what your agents may share with external agents and which commitments always require a person. Use it to evaluate any vendor agent before you enable it in your account.
Related research
Check your data before you deploy marketing agents
Agents amplify whatever data they are given, including its errors. This six-part checklist helps teams find data problems before agents spread them.
MCP for marketers: tool protocols change what agents can touch
Open protocols for connecting agents to tools make connections reusable and capabilities explicit. Marketing needs a plan for what those connections expose.
Observability for marketing agents: traces, logs, and run histories
If you cannot see what an agent did and why, you cannot improve it, defend it, or trust it. Observability is the foundation of every other agent practice.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
