# Who coordinates your company's AI agents?

Fabio Munhoz · 2026-11-20

Canonical: https://fabiomunhoz.com/blog/who-coordinates-ai-agents/

Language: en

As tools gain autonomy, the business needs to define who decides, who executes, and who is accountable for the outcome.

## In brief

Coordinating AI agents requires shared priorities and responsibilities across marketing, sales, and support. Direction sets goals and limits; orchestration distributes work; execution performs and verifies authorized tasks. Changes in an account’s situation must reach every affected workflow.

Imagine this: a customer opens a support ticket because they are struggling to use the product. At the same time, a sales agent prepares an expansion offer for that account. A third system adds the contact to a referral campaign.

Each workflow has its own rationale. Support wants to resolve the problem. Sales wants to expand revenue. Marketing wants referrals.

To the customer, all of it comes from the same company.

This is an illustrative scenario, but it raises an important question for teams adopting agents: how do you make distributed decisions respect the context of the entire relationship?

## Coordination starts with accountability

It is tempting to answer with another layer of software. A supervisor agent could receive signals, assign tasks, and resolve conflicts.

That layer can help. First, however, the company needs to decide which priorities take precedence. Who can suspend a campaign? When should a support issue prevent an offer? Which exceptions require the account owner's intervention?

Without those decisions, the supervisor is being asked to arbitrate questions leadership has not yet resolved.

Yext uses the term multiplayer agent harness to describe an environment where teams and agents share context and work together. The proposal highlights a relevant point: coordination also involves the people who set priorities and assess the work. [Source](https://www.yext.com/about/news-media/yext-announces-multiplayer-agent-harness-purpose-built-for-marketing-growth).

## A rule needs to work across tools

Return to our example. The company could establish that an open critical incident pauses expansion offers until the account owner reviews the situation.

For that rule to work, the incident needs to be linked to the correct account. The sales system needs to receive the status change. The campaign must check the restriction before sending a message. Someone must decide when outreach can resume.

The work spans data, integration, process, and authority.

I would organize coordination at three levels:

| Level | Responsibility | Example |
| --- | --- | --- |
| Direction | Choose objectives and limits | Prioritize retention during a critical issue |
| Orchestration | Assign work and resolve dependencies | Pause the offer and route the account for review |
| Execution | Perform and verify the authorized task | Update the record and confirm that sending has stopped |

People and agents can contribute at all three levels. The division depends on risk, data maturity, and the ability to verify results. The person accountable for a decision should remain identifiable even when part of the work is automated.

## Approvals need to be designed too

If every action depends on the same person, the operation ends up waiting in an approval queue. If no action requires review, exceptions can create consequences the team never intended to accept.

One approach is to set boundaries by action type. An agent can research an account and prepare a summary within an agreed scope. Changing commercial terms or publishing a product claim requires different treatment.

Those boundaries should also consider how easily an action can be corrected. A draft can be revised before it reaches the customer. A sent message has already affected the relationship, even if it is corrected later.

The system should make a proposed decision reviewable: show the evidence, the action, its expected effect, and the uncertainties. The person approving it needs to understand what they are authorizing.

## Start where processes meet

To find coordination gaps, I would bring together the people responsible for marketing, sales, support, and RevOps. Each would bring one process that already uses AI or automation.

The exercise would follow a single account through those processes: what information enters, which action can happen, who is accountable, and which event should interrupt the workflow.

Gaps often become visible at the handoffs. One system records a reply while another continues the sequence. A person corrects a record, but the next task reads an old copy. Support knows there is a problem, but the campaign never receives that signal.

In the [article on Revenue Harnesses](https://fabiomunhoz.com/blog/what-is-a-revenue-harness/), I suggested evaluating a decision from beginning to end. Coordination extends that test: we need to examine what happens when decisions from different functions meet.

**Choose an account and ask: if its situation changes today, which people and agents need to act differently tomorrow?**

The answer shows where to start building an operation that works from shared context.

## Questions and answers

### Who should coordinate a company’s AI agents?

The company must define who sets priorities and who is accountable for each decision. People and agents can participate in direction, orchestration, and execution, while the responsible human remains identifiable.

### Can a supervisor agent solve coordination on its own?

A supervisor can distribute tasks and help resolve dependencies. Leadership must first define priorities, limits, and situations requiring human intervention; software cannot independently settle unresolved organizational decisions.

### How can teams start coordinating agents across departments?

Follow one account through marketing, sales, support, and RevOps processes. Identify the inputs, possible actions, accountable people, and events that should stop or change each workflow.
