Run a one-hour agent fleet audit to find control gaps before an incident

An agent fleet audit tests containment. If you cannot name every agent, owner, and data path in one hour, the fleet is exposed. AI agents were in production at 96% of organizations. 96.4% of IT decision-makers reported complete and accurate agent inventories.

Most fleets are larger than the owner thinks

Guild.ai released The AI Agent Management Gap on 22 September 2026, using a Morning Consult survey of 362 U.S. IT decision-makers at companies with more than 100 employees, fielded from 4 to 9 August 2026. Organizations with agents ran dozens or more at a 47% rate. Organizations with agents reported an agent-related operational consequence in the previous 12 months at a 66.7% rate. Agent-related incidents hit 78% of companies with 20 to 199 engineers, versus 43% for the smallest engineering teams.

The sample is not a census, but it sets the scope for the audit.

Inventory and registration are the core agent deployment controls

Inventory completeness sets the baseline. Every agent needs a name, an owner, a purpose, and a data access list. Unregistered agents are the failure mode. IT decision-makers rated employee deployment of agents without formal approval as very or somewhat likely at a 60.6% rate. Only 35.9% of organizations required pre-deployment agent registration.

Registration makes ownership, scope, and stop authority explicit. Start with agents that can act, not agents that only read. An agent with write access to a database, a ticketing system, or a payment workflow needs a higher bar. Treat an unreachable owner as an unowned agent. Treat a missing data access list as unknown scope.

Duplicative work is a symptom of missing inventory. 61.2% of engineering and IT teams reported spending time on redundant agent efforts, including recreating or maintaining agents that were already available elsewhere.

Monitoring and audit trails turn incidents into evidence

A centralized dashboard or monitoring tool existed at 42.7% of organizations. Logging or audit trails existed at 39.8%. Without a trail, you cannot reconstruct what an agent did, which data it touched, or who approved the action.

A central view should show agent identity, recent actions, error states, and data sources. A dashboard that only shows model output is not enough. A log that cannot be filtered by agent, owner, or action type will fail during an incident. The audit trail becomes the incident response file. It should answer what ran, what it accessed, and who can stop it.

The AI agent kill switch turns the audit into action

The gap between automated and manual stopping separates quick containment from long cleanup. Just 31% of organizations had an automated kill switch that could stop a malfunctioning agent at once. The share rose to 76.5% when manual responses were included. Manual stop is not enough for a runaway agent in production. A human may be asleep, on leave, or unsure which agent to stop.

Document, test, and give the owner access to the stop path. Do not let it depend on a laptop, an account, or a person. Cover new agents created by the agent, external tool calls, and shared storage writes while preserving evidence.

Run the five-control audit in one hour. Name every agent, owner, and data access path. Time the stop path from alert to stop. Run the drill in a quiet window, not after an incident. Give each owner a fixed window to answer. Record the answer, the owner, the gap, and the time from alert to stop. The output should be a list of fixes, not a blame report.

• Inventory completeness: Can you name every agent, owner, and data access path in the drill?
• Centralized monitoring: Can you see active agents, errors, and recent actions from a central view?
• Audit trail: Can you retrieve a log that shows what an agent did and who approved it?
• Pre-deployment registration: Can you show that each agent was registered before it touched production data?
• Automated kill switch: Can you stop a named agent without opening a ticket or waiting for a human?

Fix the gaps in order of risk. Start with agents that can write, delete, pay, or send messages. Then close the monitoring gaps. Then make the stop path automatic. The audit ends when a new operator can run it without asking the original builder.