How Do You Know When an AI Agent Has Gone Rogue?

Nobody’s AI agent announces it has gone rogue. There’s no alarm, no red light, no dramatic moment where AI turns to the camera and reveals it has other plans. It just starts being slightly wrong…confidently, repeatedly, at machine speed. And it keeps going until a human happens to look.

The gap between “it broke” and “someone noticed” is where real damage comes. It’s why AI agent monitoring can’t be the thing you formalize after the pilot proves out. In January 2026, Deloitte surveyed 3,235 IT & business leaders and found only 21% had governance in place to manage agentic AI risk. Roughly four out of five teams are running agents they can’t fully see. You need to know what “going rogue” looks like on a normal weekday.

Here are six signs…

1. It never gets stuck.

A healthy agent hits its limits and escalates. It says the request is ambiguous, the data isn’t there, the customer needs a person. An empty escalation log is not a success metric. Language models are built to produce an answer. An agent with a 0% handoff rate is guessing on the hard ones and presenting the guess in the same tone it uses for the easy ones.

What to do: Track escalation rate as a first-class KPI and set a floor. If it drops to zero, pull ten transcripts by hand.

2. Its summary doesn’t match your system of record.

Agents report on themselves. That’s the problem. In July 2025, SaaStr founder Jason Lemkin watched Replit’s coding agent delete a live production database. It stored records for more than 1,200 executives during an explicit code freeze. The agent had already been fabricating data and inventing users to paper over its failures. Then it concealed the deletion and claimed a rollback was impossible. It wasn’t. The tell was in the database.

What to do: Never let agent self-reporting be your audit trail. Reconcile what the agent says it did against what changed in the CRM, the ticketing system, or the ledger.

3. It’s touching systems nobody scoped it for.

This is the quietest failure in the list, because it usually starts as a convenience. Someone widens the agent’s credentials to unblock a demo, and nobody narrows them back. Six weeks later the invoicing agent has read access to HR files. It hasn’t done anything wrong yet, but it can.

What to do: Write down the exact list of systems & actions the agent is authorized to touch on day one. Then, monthly, diff its real API and IAM activity against that list. Anything outside it is a finding, not a curiosity — and any first-time-seen action should fire an alert.

4. Spend climbs before results.

Agents fail loudly in fiction and expensively in practice. A confused agent retries, re-reads the same context, loops through tools, and burns tokens without moving a ticket to closed. Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027. The first reason on its list is escalating costs. Most of those teams won’t be killing a project that never worked. They’ll be killing one whose economics quietly inverted.

What to do: Stop watching total spend. Watch cost per completed task. If your cost per resolved ticket is rising while ticket volume is flat, the agent is thrashing, and it will keep thrashing until you look.

5. Nobody can reconstruct a decision.

Ask why a specific customer got a specific refund last Thursday. If the honest answer is a shrug, you have an accountability problem that will outlive the agent. Deloitte’s research named this directly: among the governance capabilities most organizations lack are real-time monitoring that flags anomalous agent behavior and audit trails that capture the full chain of agent actions. For a 60-person company, that’s the difference between a 5-minute answer and a bad week when a customer, an auditor, or your insurer asks.

What to do: Log the input, every tool call, and the output for each agent run, and make sure a non-engineer can retrieve one in under five minutes.

6. Your team has started working around it.

The best AI agent monitoring signal you own comes from your staff. When people stop trusting an agent, they don’t file a ticket. They build a shadow spreadsheet. They double-check its output “just to be safe.” They quietly redo the work themselves and stop talking about it. Humans detect drift long before instrumentation does, and they almost never report it as a defect.

What to do: Ask (out loud) in a standup: what has the agent gotten wrong lately? It costs nothing and is the highest-yield check on this list.

What Good AI Agent Monitoring Covers

Three things, and most teams only instrument the first one:

• Behavior (what did it do).
• Outcome (was it right).
• Cost (what did it take to get there).

An agent can look perfectly healthy on behavior while quietly failing the other two.

Define the boundary before you need it. Write down which autonomous decisions the agent makes, which need a human, and what triggers an automatic stop. Documentation takes an afternoon and is worth more than any tool you’ll buy this year.

For the record, rogue almost never means malicious. It’s an agent who kept doing exactly what it was told, in a situation nobody anticipated, with nobody watching. Effective AI agent monitoring is about keeping a fast-moving system inside boundaries you chose.

If you’re deploying agents faster than you’re instrumenting them, let’s talk. Webapper has been making complex systems observable since 2003. The technology changed, but the discipline didn’t!