AI teams agreed on the wrong answer in 75% of debates — and lower token bills hid the failure

Presented by Outshift

Enterprises that give AI agents shared memory and carefully tuned context still end up with agent teams that fail at shared goals, says Vijoy Pandey, senior vice president and general manager of Outshift by Cisco, the company's incubation arm. In fact, a recent Berkeley study found that roughly a third of the agent failures it examined could be attributed to misaligned goals.

Much of the problem is that agents loaded with all the institutional knowledge a company can supply still pursue the goals of their own role, Pandey told VB CEO Matt Marshall during VentureBeat’s Agentic Infrastructure Stack Summit. Put a sales agent, an engineering agent, a product agent, a finance agent, and a legal agent together to close a deal, and each one optimizes for what its own function rewards.

"Do you think you'll get the outcome the enterprise wants from that deal? The answer is no," Pandey said. "These agents are optimizing for their own local incentives, so they don't align on intent, they don't coordinate or negotiate among themselves, and once they finish a job, they don't verify as a team what they've accomplished."

Much current thinking is that the strongest AI teams spend fewer tokens on fewer agents, but this thinking is incomplete in two ways, Pandey said. Outshift's research shows that a low token bill can mean the agents skipped the work of reaching a correct agreement, and as enterprises run more agents over the next few years, every new vendor and specialist they add multiplies the agreements those agents have to reach.

Outshift found five-agent AI teams agreed on the wrong outcome in 75% of debates

Outshift has spent eight to nine months studying multi-agent systems, including deployments it runs in production for healthcare, video surveillance, and platform engineering. In one category of failure, every session the team examined showed cognitive failures, or the agents failing to align on a goal, coordinate their work, negotiate tradeoffs, or check the result.

When Outshift staged debates among five-agent teams, 75% ended with every agent agreeing on the wrong outcome. The agents backed off their positions and converged on a consensus, and a cost report would have logged those sessions as efficient.

"Yes, they were spending fewer tokens, because they weren't doing the hard work of aligning, coordinating and verifying what they were supposed to do," Pandey said.

That means cost per agent isn't the best measure of success. A team that debates a question till it reaches resolution burns more tokens than a team that folds early, so a falling bill can mean the agents dropped the disagreement that would have caught the error.

“Agents are like experts on billable hours. In fact, they’re billing by the word, by the thought,” Pandey said. “You want to get the outcomes of a team, and not just pay for five experts’ thoughts.”

Standard monitoring misses the problem too. Observability tools track latency, errors and, increasingly, token spend, and a team that converges on a wrong answer trips none of those alarms.

"All of your dashboards from every observability vendor will show green, while your outcomes are going red," Pandey explained.

Shared context isn't enough to align AI agents on a goal

Asked whether shared memory could solve the problem, Pandey pointed to what it leaves unfinished. Marshall cited PayPal and other companies discussed earlier in the day that are building memory systems with standard operating procedures and other shared knowledge for agents. While enterprises need that kind of team memory, Pandey said, they also need coordination mechanisms: one for agents to agree on intent before work begins, one to coordinate and negotiate during the work, and one to verify the global outcome, not just each agent’s local incentives.

Things break without them. A sales agent might learn from a customer that a deal hinges on a feature supporting 100,000 users and relay that requirement to the engineering agents. Engineering builds for 100,000 cumulative users, while the customer meant 100,000 concurrent users. That's a goal alignment failure, and it happens before any agent starts work.

Once work begins, each agent's loop optimizes for its own function's metric. The sales agent pushes for the fastest revenue, the engineering and product agents cycle through feature, test, and deploy, the legal agent checks contractual obligations, and the finance agent protects margin. Every one of them can hold the best institutional knowledge for its role while the deal stalls.

"That's a coordination problem," Pandey said. "You cannot solve coordination through context alone."

Enterprises will run more AI agents for specialization, multiple vendors, and cognition engines

The coordination problem will continue to grow, because the number of enterprise agents keeps growing for a number of reasons. For instance, hyperspecialization. Companies build agents for narrow tasks, give each one the context its team needs, and grant it only the permissions that team holds.

"You can't give the finance agent the same access engineering has. It doesn't work," Pandey said.

Specialization will grow inside individual agents as well, with conversational models handing work to back-end models tuned for reasoning and for taking action.

Heterogeneity will also grow the number of enterprise agents. A single deal desk workflow might involve agents from Salesforce, Cisco, Microsoft, Harvey, and Workday, each built on its vendor's own stack.

The third kind of driver is what Outshift calls cognition engines, agents that IT teams deploy alongside business workflows. Guardrail engines handle security, scale, compliance, and availability, while accelerator engines handle work such as memory distillation and joint reasoning across agents. An enterprise's business logic agents will have to cooperate with both kinds, and all three trends point to more agents per company over time.

Finding where agent teams go wrong

Outshift spent the past 12 to 18 months connecting agents through its open-source AGNTCY project, part of the Linux Foundation. Running those systems in production exposed the next problem: agents could communicate without working toward the same outcome.

The first step is making that failure visible. Outshift’s newly announced Scope examines the OpenTelemetry traces enterprises already collect, looking at what agents say and do to detect cognitive drift from the shared goal. It builds on semantic extensions to OpenTelemetry that Outshift introduced with Microsoft last year.

What happens next depends on where the agents lose alignment. If they start with different interpretations of the task, Outshift’s Accord protocol gets them to agree on intent before work begins. If they diverge during execution, Concord supports coordination and negotiation. Both work above Agent2Agent (A2A) and Model Context Protocol (MCP) standards. Agreement still needs to produce the right result. That is the role of MAS Lab, which tests whether a multi-agent system delivers the outcome its owners specified. Together, the tools extend monitoring beyond whether individual agents ran successfully to whether their work added up to a successful team.

Outshift published all four at github.com/iocproject, as the start of what it calls the Internet of Cognition, a layer above agent connectivity where agents align, coordinate, and verify their work, which Outshift is building in the open. The tools aim to make agreement something enterprises can test against the intended outcome. As Pandey explained later in the session, consensus itself can be the failure.

“Yeah, this is groupthink,” he said. “Nobody's challenging, nobody is arguing, and everybody's accepting the other agents’ outcomes as ground.”