← Research library

Roles and permissions for a mixed human and agent team: name who can act, approve, and stop

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.

Marketing teams already manage permissions: who can publish to the website, launch campaigns, change budgets, or export contacts. Agents add a new kind of team member to that model, and they need permissions of their own.

The people who manage, review, and override those agents need defined permissions too. Without a clear model, access drifts. Agents inherit broad rights from whoever set them up, and after a few months nobody is sure who is allowed to change what.

How permissions drift

Drift rarely starts with a bad decision. It starts with a convenient one. Someone connects a new content agent to the CMS using their own login because the integration account is not ready yet. The agent works, the deadline passes, and the shortcut stays.

A month later the same agent needs to read campaign results, so it gets read access to the analytics tool. Then it needs to update a field in the CRM. Each grant is small and reasonable on its own. Together they give one agent the reach of a senior marketer, with none of the review that a senior marketer's access would get.

The second source of drift is people changing roles. The person who approved an agent's permissions moves to another team, and nobody inherits the decision. The permission outlives its reason. In our research, we treat both patterns as the default outcome when nobody owns the model, rather than as rare mistakes.

Roles to define

We find five roles cover most programs. Each one answers a question that otherwise gets answered by whoever happens to be online.

  • Agent owner: responsible for an agent's performance and behavior, can change its instructions, and requests permission changes. Accountable for its outcomes.
  • Reviewer: approves or rejects agent outputs within a defined area, but cannot change the agent itself.
  • Operator: monitors agent runs, handles escalations, and can pause an agent. Often part of an agent operations function.
  • Permission approver: approves changes to what agents can access and do, and should be someone other than the agent owner for high-risk permissions.
  • Administrator: manages the underlying systems and the integration identities agents use.

In small teams, one person may hold several roles. That is fine. The point is to know which role they are acting in at a given moment, so the record shows whether a change was an owner's request or an approver's decision.

The split between owner and permission approver matters most. An owner naturally wants their agent to do more. A second person reviewing high-risk grants means no one can give themselves, or their agent, broad capabilities without another look.

Agent permissions

Each agent should have its own identity in every system it uses, never a person's credentials. A separate identity makes the agent's actions visible in logs and lets you revoke its access without locking out a colleague. It also stops the agent from carrying a person's broad rights into a narrow task. Our post on connecting agents to your CRM covers how this works for the system most agents touch first.

Beyond identity, we recommend four principles for every grant.

  • Least privilege: grant only the access the agent's current tasks require.
  • Explicit grants: document every permission with a reason and an approver.
  • Time limits: set temporary permissions to expire automatically.
  • Regular review: check each quarter that permissions still match actual use.

Permissions should also match the agent's level on the autonomy ladder. An agent at Level 2, where a person approves each action, needs the right to prepare actions but not to execute them alone. An agent at Level 3 needs execution rights, but only within the limits written into its tools. When an agent moves up or down the ladder, its permissions should change in the same step.

Separation of duties

Some combinations of capability should not sit with a single agent or person without review. These are the pairs we check first.

  • Content: creating content and approving it for publication.
  • Audiences: building an audience and sending to it, above a size the team sets.
  • Budgets: changing a budget and approving the change.
  • Access: granting permissions and using them.

These separations echo long-standing financial and security controls, and they work for the same reasons. A single actor holding both halves can make a mistake that nobody sees until it reaches customers.

In an agent program, the cleanest place to enforce separation is at the tool layer. In our Agentic Hub architecture, agents reach external systems only through governed connectors with scoped permissions. If the content agent's connector has no publish action, no instruction can make it publish.

A worked example

Consider a hypothetical B2B software company running four agents: research, content, lifecycle email, and paid media. The marketing operations lead decides to map every role and permission before adding a fifth agent.

The audit turns up three problems. The content agent publishes to the blog using a product marketer's login, so its posts appear under her name in the CMS history. The lifecycle agent can both build segments and send to them, with no size limit. And the paid media agent has a budget permission that the person who approved it, now on another team, granted for a launch that ended months ago.

The fixes follow the model. The content agent gets its own identity with draft rights only, and publishing moves to a reviewer. Segments above a set size now need a person's approval before any send. The expired budget permission is removed, and the paid media agent's owner files a new request with a reason and an expiry date.

None of this required new software. It required writing down who holds each role and checking the grants against it.

The pause button

Every agent should have a simple way to be paused immediately, and the people who can press it should be named. When something goes wrong at an inconvenient hour, nobody should be searching for who has access.

We suggest at least two named people per agent, in different time zones if the team spans them. The pause should stop new actions at the connectors, not just ask the agent to stop. Test it on a schedule, the same way you would test a backup, so you know it works before you need it.

What to measure

A permission model is only as good as its upkeep. We recommend tracking a few signals each quarter.

  • Unowned agents: agents with no named owner, operator, or pause holder.
  • Grants without reasons: permissions with no documented reason or approver.
  • Stale grants: permissions unused for a full review period.
  • Pause test results: whether each agent stopped when tested, and how long it took.

These fit naturally into a quarterly governance review. The goal is a short list that gets shorter each quarter.

What we are studying next

The harder questions start when agents hand work to other agents. If a router agent assigns a task to a specialist, which agent's permissions apply, and who approved the combination? We are studying how permission checks should travel with a task across handoffs, so a chain of agents never holds more access than any single grant allowed.

How to apply this: for each agent, write down its owner, its reviewers, who can pause it, and who approved its current permissions. If any of these is unclear, or an agent still uses a person's credentials, fix it before expanding the agent's scope.

Related research

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.