---
title: "AI Agent Governance: Who Owns Each Agent and How to Switch It Off in Minutes"
description: "AI agent governance in 7 controls: inventory, owner, permissions, traces, evals and a kill switch. Supervise your agents without creating another committee."
url: https://addabra.com/en/blog/ai-agent-governance
lang: en
date: 2026-11-09
category: "Harness engineering"
keyword: "AI agent governance"
publisher: Addabra
alternate: https://addabra.com/blog/gobernanza-de-agentes-de-ia
---

# AI Agent Governance: Who Owns Each Agent and How to Switch It Off in Minutes

> AI agent governance in 7 controls: inventory, owner, permissions, traces, evals and a kill switch. Supervise your agents without creating another committee.

Friday, 10:40 p.m. A customer writes in, upset: they just received their third collections email in an hour, for an invoice they paid last week. The on-call manager looks into it and finds a collections agent resending reminders in a loop. He asks in the team chat who set it up. Nobody answers. Nobody knows whether it can be paused without breaking billing, or who has permission to do it.

The agent is not malicious. It is simply ungoverned. This article explains **AI agent governance** as what it really is: a handful of operational controls that tell you which agents you have, who is accountable for each one and how to switch it off in minutes.

## Why it happens: agents in production, governance up in the air

Agents arrive at companies through many doors. One department buys a tool, another builds its own workflow, a vendor connects an assistant to the CRM. Each one solves a real problem, which is why they stay. What almost never comes with them is a plan to supervise them.

A March 2026 survey of 650 technology leaders found that 78% have at least one agent pilot, but only 14% have scaled it across the organization. That gap is not explained by technology alone. Scaling means being able to answer uncomfortable questions: who owns this agent? What data does it see? What happens if it gets something wrong at three in the morning?

When those questions have no answer, three symptoms show up:

- **Agents with no owner.** Whoever set it up moved to another team and nobody inherited the responsibility.
- **Too many permissions.** The agent can write to systems it only needed to read.
- **No brakes.** Stopping it requires calling an engineer, and the engineer does not pick up on a Friday.

## AI agent governance: controls, not committees

"Governance" tends to bring to mind committees, forty-page policies and monthly meetings. Those have their place, but they do not stop a looping agent on a Friday night. Governance that works lives inside the system, in the *harness*: the software layer around the model that defines what it can see, what it can do and who supervises it.

Think of it as an aircraft's instrument panel. It does not replace the pilot, but it shows what is happening at every moment and puts the controls within reach. These are the seven controls we recommend, with the question each one answers and the evidence you should be able to show.

| Control | Question it answers | Evidence you should be able to show |
|---|---|---|
| 1. Agent inventory | Which agents do we have and what is each one for? | A registry with purpose, owner, data accessed, tools, model and monthly cost cap |
| 2. Accountable owner | Who answers for this agent? | A named person and a backup, per agent |
| 3. Least privilege and approval by role | What can it do alone and what needs sign-off? | Permission list per agent and an approval log |
| 4. Traces and audit log | What did it do, when and with what information? | A trace of every action, searchable by date, customer or agent |
| 5. Evals before every change | How do we know the change did not make things worse? | Test suite results before release |
| 6. Kill switch and degraded mode | How do we stop it and who picks up the work? | A tested procedure, with response time measured in minutes |
| 7. Periodic review | Does it still make sense as it is? | Short review notes with the changes decided |

## The seven controls, one by one

### 1. A living inventory
Every agent has a record: which process it handles, who owns it, what data it can see, which tools it can use (send emails, create invoices, query the CRM), which model it runs on and how much it can spend per month. If an agent is not in the inventory, it should not be in production.

### 2. A named owner
Not "the IT team," but a business person who understands the process and is accountable for the agent's results. With a backup for vacations and weekends. They are the first call when something goes wrong.

### 3. Least privilege and approval by role
The agent gets only the permissions its task requires, and consequential actions (payments, customer commitments, irreversible changes) wait for sign-off from a defined role. We go deeper in [human approval by role](https://addabra.com/en/blog/human-approval-by-role) and [AI agent security](https://addabra.com/en/blog/ai-agent-security).

### 4. Traces of everything
Every decision is recorded: what the agent received, what it looked up, what it did and how it responded. Without traces, an audit is a reconstruction in the dark. More detail in [observability for AI agents](https://addabra.com/en/blog/ai-agent-observability).

### 5. Evals before release
An *eval* is an automated test that measures whether the agent does its job well on real cases. Changing the prompt, the model or a tool without running evals is shipping blind.

### 6. Kill switch and degraded mode
Someone on call, without writing code, can pause the agent in minutes. And the process does not stop: tasks move to a queue handled by people, with the context the agent had already gathered. This is the control most often forgotten, and the one that matters most on a Friday night.

### 7. Periodic review
Every quarter, the owner reviews cost, quality, incidents and permissions. Some agents get expanded, others get trimmed and some get retired.

## Regulation: what to keep on your radar

If you sell in Europe, the EU AI Act entered into force in August 2024 and applies in phases. Its transparency obligations (Article 50), which include telling people when they are interacting with an AI system, apply from August 2, 2026. Other obligations follow their own timelines, which have been subject to changes.

This is not legal advice: confirm with your legal counsel which obligations apply to you and from when. What we can tell you is that the seven controls above put you in a better position under any regulatory framework, because they produce evidence. And they pay off right away: enterprise customers' security questionnaires and audits get answered in hours, not weeks.

## Illustrative example: the same Friday, with governance

Back to the collections agent.

**Before:** the agent is not in any inventory. It has permission to send unlimited emails. There are no searchable traces. The on-call manager does not know whether switching it off will break billing, so he waits until Monday. The customer receives eleven emails.

**After:** the manager opens the inventory and sees the agent's owner is the head of collections, with an on-call backup. He hits pause; the agent stops and pending reminders move to the human team's queue. The trace shows the agent did not recognize the payment because it arrived with a different reference number. On Monday, the owner fixes the rule, the evals confirm the case is now handled correctly and the agent goes back into production. The customer received two emails and a timely apology.

## Common mistakes

1. **Confusing governance with a document.** A policy nobody can act on at 10 p.m. is not a control.
2. **Assigning ownership to a department, not a person.** "Operations" does not answer the phone.
3. **Having a kill switch with no degraded mode.** If switching off the agent stops the process, nobody will dare to use it.
4. **Granting broad permissions "so we don't slow things down."** Every extra permission is a potential incident.
5. **Never testing the shutdown.** A procedure that was never rehearsed fails exactly when you need it.

## What to do on Monday

List every agent running in your company today, including the ones a vendor brought in. For each one, write down its owner, its permissions and how it gets switched off. The blank cells are your work plan.

At Addabra, we build every agentic system with these controls from day one: an inventory, an owner per agent, least privilege, approval by role, traces, evals and a kill switch with a degraded mode tested before go-live. If you already have agents running and want to bring them under control, we can help you [take your AI to production](https://addabra.com/ai-to-production/en) with governance built into the harness, not bolted on top.

## Frequently asked questions

### What is AI agent governance?

It is the set of controls that tells you which agents you have, who is accountable for each one, what they are allowed to do and how to stop them. It is not a committee or a policy document: it is concrete mechanisms inside the system, such as an inventory, least-privilege permissions, traces and a kill switch.

### Do I need an AI committee to govern my agents?

Not to get started. What you need is a named owner for each agent, an up-to-date inventory and controls that work without waiting for a meeting. A committee can review the whole picture every quarter, but day-to-day governance has to live in the system.

### What is a degraded mode for an AI agent?

It is the plan that keeps the process running when the agent is switched off: tasks go back to a queue handled by people, with the context the agent had already gathered. That way, switching off an agent does not mean stopping the operation.

## About Addabra

Addabra is an AI-software-first software factory that designs and builds custom agentic systems for companies: one agent per process, a shared business memory, orchestration, human approval by role, auditing and cost control. Free diagnosis: https://addabra.com/en/blog/ai-agent-governance#contacto