AI Agent Teams Agreed on the Wrong Answer 75% of the Time While Monitoring Showed No Problems

Quick Facts

• In Outshift by Cisco's internal tests, 75% of five-agent debates ended with every agent agreeing on the wrong answer.
• Across 880 debates analyzed, accuracy dropped from 99% to 16% when cognitive failures occurred, yet standard monitoring tools flagged nothing.
• Adding a coordination protocol raised decision success rates from 33% to 93% across 14 test scenarios.

Cisco's incubation arm, Outshift, spent eight to nine months studying multi-agent AI systems in production and found a category of failure that standard tooling cannot detect. As the research shows, agents can reach unanimous agreement, report success, and be wrong — all without triggering a single latency alert or error log.

Vijoy Pandey, SVP and GM of Outshift, summarized the gap plainly: "The bottom line is, we can send messages, but agents do not understand each other, so there is no grounding, negotiation or coordination or common intent."

The Shape of the Failure

Outshift runs multi-agent deployments in healthcare, video surveillance, and platform engineering. In every session the team examined, agents hit at least one coordination failure. The problem is not communication — agents pass messages — it is cognition. Each agent interprets goals independently, and errors introduced early compound as they pass downstream.

The consensus mechanism makes things worse, not better. When researchers had groups of large language model agents debate and vote, correct-to-incorrect answer flips outnumbered the reverse. On MMLU benchmarks, groups of mixed-capability agents saw accuracy drop 8 to 12 percentage points after debate rounds. Among questions where agents initially disagreed, 23.9% converged to unanimous wrong consensus by round three.

Confidence scores offer no warning either. In debates that ended with the right answer, confidence rose an average of 5.54 points. In debates that ended wrong, it rose 10.49 points — nearly twice as far.

Why Monitoring Misses It

Observability tools track latency, errors, and token spend. A team that talks itself into the wrong answer trips none of those alarms. The architectural gap means lower token bills can actually signal a failure: agents converging fast on a wrong answer consume fewer tokens than agents that surface and resolve genuine disagreement.

Pandey described the broader coordination problem facing companies that break work into smaller AI teams: "What's happening is that since the productivity unit is shrinking, the coordination surface and the coordination tax is exploding. Smaller teams means intra-team coordination is tending towards zero, but inter-team coordination has just exponentially exploded, because now we have an N squared problem."

Outshift's Open-Source Response

Outshift has released four open-source tools under the Linux Foundation's AGNTCY project to address what it calls silent cognitive failures.

• SCOPE (System for Cognitive Observability, Provenance and Explainability): Built on OpenTelemetry, it checks intent alignment before work begins, monitors coordination during execution, and evaluates outcomes after.
• ACCORD: A protocol that forces agents to agree on a shared interpretation of the task before any work starts.
• CONCORD: Handles coordination and negotiation when agents diverge during execution.
• MAS-Lab: A testing environment that checks whether a multi-agent system actually delivers the outcome its operators specified.

Internal testing showed the difference a coordination layer makes. Without structure, uncoordinated agent groups reached a correct decision about 33% of the time. With a coordination protocol in place, that figure rose to 93% across 14 scenarios.

What This Means for Builders

The findings put pressure on teams deploying multi-agent pipelines in production. A system that returns confident, unanimous output is not a reliable signal of correctness. One cited study found failure rates between 41% and 87% across evaluated multi-agent systems, which Pandey frames as an architectural problem, not a prompting problem.

The coordination layer Outshift describes adds overhead, but Pandey argues the alternative is worse: agents that agree, report success, and are wrong, with no alert fired and no log entry written. For companies using multi-agent systems in regulated industries or high-stakes decisions, that failure mode is not theoretical. Outshift's own production deployments in healthcare surfaced it repeatedly.