A green agent dashboard can hide a task that was never finished

Vijoy Pandey, who runs Outshift by Cisco, argued on The Cognitive Revolution that a system of AI agents can look healthy on a dashboard while failing at its job. Uptime tells you the agents are running. It does not tell you the work got done.

What was said

Pandey described a test of a multi-agent system on a simple consumer job: booking a travel itinerary. He said today's dashboards only show whether agents are up and working. By that measure, the system was 66% up, which was defined as working perfectly.

Task completion was 0%. The agents had, in his words, complete misalignment on what they were trying to achieve. His summary: "your dashboards might be green, but your outcomes might be wrong, completely wrong."

He called these "silent failures with really loud consequences". Cisco's answer is Scope, an open-source tool it describes as cognitive observability. It sits on top of OpenTelemetry traces, the open standard that most applications already use to record what they do. Its aim is to show where agents misunderstand each other.

Why it matters

Our reading: a green dashboard moves the risk without removing it. If the screen says all is well, nobody checks the result. The person who finds the failure is the customer with no booking, or the small business whose order never went through.

For managers and buyers, the practical step is a question about measurement. When a vendor shows agent uptime, ask what it counts. Is it agents running, or tasks finished correctly? Pandey's example suggests those two numbers can sit far apart.

It also shifts who has to do the checking. A team that runs agents on real work needs someone who judges outcomes, because a status light will not. That judgment is a job, and it needs time and skill.

The other side

The figures come from one simple example that Pandey described. The excerpt does not say how many runs were tested or how the setup was built, so it is hard to know how often this gap appears in real deployments.

Pandey also has a stake in the answer. Cisco is releasing Scope, and he urged listeners to use it. Nothing in the excerpt shows independent evidence that it catches failures in live systems.

Pandey also did not discuss who pays when an agent fails silently. That part is our inference from his warning about loud consequences.