An AI agent control plane is the part of a company's systems where rules for AI agents are set and enforced: which identity each agent acts as, what it may access, which actions need a person's approval, how access is revoked, and what record is kept. It separates deciding what agents may do from the agents doing the work.
Where the term comes from
In networking and cloud infrastructure, the data plane moves the traffic and the control plane decides the rules it follows. Keeping the two apart means rules can be set, changed and audited in one place.
Applied to AI agents, the data plane is the agents doing work: reading records, drafting messages, updating tools. The control plane is where the company decides and enforces what they may do.
What an agent control plane covers
- Identity and delegation: which person, or which narrowly scoped service identity, each agent acts for.
- Access: which sources and records the agent may read, limited to what the task needs.
- Actions: which tools the agent may use and which actions it may take on its own.
- Approval: which actions wait for a person, based on risk. External messages and consequential internal changes are common candidates.
- Revocation and expiry: how access is withdrawn, and whether it is checked again at the moment an action is taken.
- Records: an audit trail of approvals and actions, including those that failed or had an unclear result.
Why enforcement belongs outside the model
A rule written only into an agent's instructions is a request, not a control. Inputs the agent reads, such as an email, a web page or a document, can contain text designed to change its behaviour. This is prompt injection.
A control plane enforces rules in the surrounding system: the permissions check happens before data is returned, and the approval check happens before an action is sent, whatever the model says. Treat retrieved content as untrusted, keep agent permissions narrow, and review proposed actions and their details before approving them.
Why one control plane beats rules in every tool
When each AI tool carries its own settings, a company ends up with rules that differ by tool, nobody with the full picture, and no single record to review. A control plane that covers a company's agents gives one place to answer the questions a customer, an investor or the team will ask.
- Which agents can reach which data?
- Who approved this action, and when?
- What happens to an agent's access when a person leaves or an account is disconnected?
- Which actions failed or are still waiting?
Getting started
- List every agent and automation, and the identity each one acts as.
- Replace shared admin connections with narrowly scoped access.
- Decide which actions need approval, starting with anything sent outside the company.
- Make sure approvals and actions are recorded somewhere a change after the fact would show.
- Review the record and each agent's access on a regular schedule.
The guide to AI agent governance covers the policy side, and AI audit trails covers the record.
About Hoppflow
Where Hoppflow fits
Hoppflow provides these scoped controls for the agents and automations in its workspace, alongside a company's own policies and security measures.
- Roles define who can see, run and approve supported parts of a workspace, and supported company-brain retrieval paths apply each person's access rules.
- An agent's draft to a customer, supplier or connected tool is held for a person's approval, and Hoppflow checks access again before a supported external action is sent.
- Approvals and external agent actions are kept in a tamper-evident authorization record.
- Hoppflow records a held, failed, or unclear external effect as such instead of filing the work as done.
- Disconnecting a supported integration revokes Hoppflow's ongoing access to that account.
The full set of controls is on Security and trust, and access is through the waitlist.
Questions people ask
What is an AI agent control plane?
It is where a company sets and enforces rules for its AI agents: which identity each agent acts as, what it may access, which actions need a person's approval, how access is revoked, and what record is kept of approvals and actions.
What is the difference between a control plane and a data plane for AI agents?
The data plane is the agents doing work, such as reading records and drafting messages. The control plane decides and enforces what they may do. Keeping them separate lets rules be set, changed and audited in one place.
Can a system prompt act as a control plane?
Not on its own. Instructions inside a prompt can be overridden by inputs such as prompt injection. Controls such as permission checks and approval gates should be enforced in the surrounding system, before data is returned or an action is sent.
Does governance for AI agents need shared company memory?
In a business, usually yes. Rules decide what an agent may do, but the agent still acts on whatever context it reads. Shared, current company memory, together with a workspace where the work runs, is what a governed AI workspace combines.
Does human approval make AI agents safe?
It helps, especially for consequential or external actions, but it is not sufficient on its own. Reviewers can miss problems. Combine approval with least privilege, controls enforced outside the model, testing and a reviewed record.