Cisco's Pandey: three unaddressed misalignments cut five-agent accuracy to 16%

Vijoy Pandey, who leads Outshift by Cisco, argued on The Cognitive Revolution that mistakes of meaning between AI agents can cut accuracy sharply. In simple five-agent deployments his team looked at, a few unaddressed misalignments took accuracy from near perfect to close to random guessing. Each extra failure cost far more than the one before.

What was said

In plain terms, a semantic misalignment is a case where agents misread each other's intent as work passes between them. Pandey said that in five-agent systems where this was not handled, one misalignment dropped accuracy from 99% to 98%. Two dropped it to 81%. Three dropped it to 16%, which he described as "getting close to just randomly picking an outcome".

His explanation is that the value sits in the handoffs, not in any one agent. He gave a sales example. An agent's estimate can become a commitment, then contract language, then a marketing claim. Yet the evals in place, he said, scored each agent on its own local task.

He also said public traces from ChatDev, a multi-agent development environment, showed semantic failures in 100% of sessions. In a travel-booking test, the system counted as 66% up, yet task completion was 0%. He summed it up: "your dashboards might be green, but your outcomes might be wrong, completely wrong."

Why it matters

Our reading: the usual health check, whether every agent is running, can say nothing about whether the work got done. Anyone building or buying a multi-agent workflow should ask for tests on the end result across handoffs, not just scores per agent.

Pandey's deal-desk example illustrates how a silent error could travel. A loose estimate could harden into a contract. He offered it as an illustration of the chain, not as a measured loss.

The other side

The figures come from Cisco's own deployments, and Pandey is promoting Scope, infrastructure the company is open sourcing to make these failures visible. The excerpts give no sample size and no definition of accuracy. Treat the numbers as one team's report, not an industry benchmark.

The claim is also narrower than a simple "more agents, more risk". The steep drops came when misalignments went unaddressed in five-agent systems. Pandey gave the four figures but did not explain why they fall that way, and he did not say that adding agents alone causes them. He even said an 81% result was one he could maybe live with, so the tolerable level depends on the task.