When Connected Agents Still Disagree: Vijoy Pandey on Meaning, Handoffs, and Accountability

1. Guest Background

Vijoy Pandey joins The Cognitive Revolution as senior vice president and general manager of Outshift by Cisco. The supplied guest background describes Outshift as a Cisco team involved in deploying multi-agent systems and developing and open sourcing protocols for agent collaboration. That gives this interview a particular perspective: Pandey is discussing problems encountered through deployment work while also explaining the infrastructure his team wants others to adopt.

He reports work with Swisscom, ServiceNow, SoftServe, and Webex across network configuration, cloud platform engineering, and video surveillance. These are relevant experience claims, rather than a substitute for evidence about each system's effectiveness. The interview offers selected accounts of that work; it does not present a customer-by-customer evaluation or independently verified performance record.

In this approximately ninety-minute indexed episode, Nathan Labenz presses Pandey on the gap between connecting agents and making them understand one another. The discussion covers enterprise workflows, semantic observability, alignment and negotiation protocols, personal assistants, and the supervision problems posed by higher-dimensional communication. Its central subject is therefore the organization of shared work: how specialists preserve meaning across boundaries, recognize disagreements, and remain answerable to an overall objective. Reading Pandey as both a practitioner and an advocate helps distinguish reported experience from proposed architecture and future ambition.

2. What the Episode Covers

Pandey organizes recent agent development into three stages: isolated agents, islands of agents working within a common harness, and connected agents crossing vendors, cloud providers, and organizations. The final stage expands the available division of labor. It also brings separately developed assumptions and objectives into the same workflow. The interview asks what must happen after those participants can already exchange messages.

Swisscom supplies the most concrete infrastructure example. In Pandey's account, the team selected data-center migrations and configuration changes as a relatively contained starting point. Data centers were still a large problem, but their redundancy offered better protection against failure than the mobile-network alternative he discussed. The implementation combines five agents exploring possible configurations with a formal representation of the environment that prevents prohibited operations. Pandey describes this as a neurosymbolic approach and claims a 100% guarantee of configuration success. The episode does not specify the complete conditions behind that guarantee.

A multimillion-dollar enterprise deal provides a different view of specialization. Sales interprets the customer request; legal handles contractual terms; finance checks pricing, margins, and discounts; product and engineering determine feasible features and delivery; marketing considers public claims. A coordinating verifier must keep those contributions consistent. Pandey says that verifier is usually human today. The example explains why agents from different suppliers might need to participate, while shifting attention toward the dependencies between their outputs.

The proposed infrastructure has three named components. Scope is presented as an open-source cognitive observability layer built on standard OpenTelemetry traces. Accord addresses intent and terminology before work begins. Concord supports coordination and negotiation while work proceeds. According to Pandey, the latter two are released as skills for existing agent environments and can also operate through Agent Bridge above A2A and MCP. The intended arrangement adds alignment mechanisms to existing communication rather than requiring a single shared model or runtime.

The episode also marks important differences in readiness. Pandey explicitly calls the Webex collaboration space under discussion a demonstration, not a production deployment. A third protocol for checking finished work against agreed objectives had not been released. Those statements prevent the overall vision from being mistaken for a completed, uniformly deployed stack.

Personal agents bring the discussion back to everyday friction. Pandey describes his Arnold and his wife's Lyra attempting to arrange a celebration dinner two weeks ahead. Model limitations and repeated requests for permission to send messages obstruct progress, but so do disagreements about food, location, ambience, menus, and availability. The example is useful because removing one technical obstacle would not eliminate the negotiation. Even a modest shared task requires the agents to reconcile distinct preferences and recognize an acceptable joint outcome.

3. Core Views: Reasoning, Examples, and Limits

The strongest part of Pandey's argument concerns what changes during a handoff. An engineering estimate can become a sales commitment, then contractual language, then a marketing claim. Each participant may produce an apparently competent local output while increasing the certainty or obligation attached to the original information. Evaluating those participants separately can miss the unauthorized transformation between them. The relevant unit of correctness is the chain of commitments and its enterprise outcome, including the qualifications that should survive each transfer.

His example of 100,000 users makes the mechanism tangible. Engineering understands a total user population; the customer means simultaneous users. Monthly or daily active users introduce further possible meanings. Repeating the same number does not establish agreement about the task. Alignment requires a shared referent and conditions under which the requirement will be considered satisfied. A delivery date creates a related problem: one participant may mean an alpha, another a first release, another final delivery, and finance a contract or payment milestone. As customer feedback and scope evolve, an initially shared interpretation can become obsolete.

That is the reasoning behind a lifecycle of alignment before work, negotiation during work, and verification afterward. In Pandey's account, Scope monitors drift and can trigger Accord or Concord to restore agreement. This is a coherent specification of what the system should manage. It is not, by itself, evidence that the detector reliably identifies those conditions. His lane-correction analogy clarifies the intended intervention, but the conversation does not establish detection coverage, false alarms, or the handling of disagreements that remain unresolved.

The reported performance figures should be read with the same distinction. Pandey says every ChatDev session his team examined contained semantic or cognitive failures. He also describes some five-agent deployments in which accuracy fell from 99% to 98% and then 81% as failures accumulated. In a travel-booking example, he reports 66% apparent operational uptime alongside 0% task completion. Without sample sizes, complete metric definitions, or experimental controls, these are not universal failure rates or a general degradation curve. They nevertheless identify a consequential evaluation gap: operational availability, shared meaning, and successful completion can diverge sharply.

Swisscom illustrates a complementary response to uncertainty. Agents explore possibilities while a formal model blocks actions that violate represented rules. This division can place explicit constraints around generative exploration. However, the meaning of Pandey's 100% guarantee depends on what the formal representation includes and what assumptions remain true in the environment. The interview does not establish those boundaries or provide independent validation. Preventing violations recognized by a model is a narrower proposition than showing that every relevant real-world failure has been excluded.

Pandey's architectural preference follows from his expectation of persistent heterogeneity. He describes harnesses as both accelerators and sources of guardrails, and argues for separating them from models while using open standards across organizations, clouds, identities, and access arrangements. That is a rationale for interoperability work, not proof that general interoperability has been achieved. His response to the possibility that stronger models will absorb today's scaffolding similarly establishes a position rather than a scaling result. Vertical capability, algorithmic efficiency, and horizontal collaboration can develop together; the interview does not confirm Labenz's suggested square-root relationship between agent count and speedup.

The final exchange exposes a deeper boundary. Labenz worries that transferring high-dimensional internal states would make supervision harder when readable reasoning is already difficult to monitor. Pandey supports investigating latent-space protocols for demanding problems such as drug discovery, materials, logistics, financial modeling, and weather modeling. His response to the prospect of incomprehensible AI superstructures is that people already depend on experts to understand complex trading and legal systems. He expects specialist-agent builders to understand their own systems. That expectation is a premise of his answer, not something the interview demonstrates will remain true as capabilities increase.

His proposed system of record has practical value within that uncertainty. Recording actions, decisions involving other agents, and tool calls could support compliance checks and fault investigation, with corrective mechanisms responding to detected drift. But tracing external behavior and interpreting internal high-dimensional exchanges are different capabilities. The episode provides no concrete monitoring method or quantitative validation connecting the two. A ledger may help investigators reconstruct what happened and locate a disputed handoff; it does not automatically reveal why agents reached an internal agreement or whether humans can meaningfully supervise that agreement.

4. Learning and Application

A practical application of the interview is to make the handoff itself reviewable. In a deal workflow, preserve the original requirement alongside each specialist's interpretation and its approval status. Specify whether a user target concerns total or concurrent users, and attach each date to a delivery stage. Before an estimate enters a contract, require an explicit decision about whether it has become an approved commitment. These are implementation suggestions derived from Pandey's examples, not procedures shown to have produced customer results in the episode.

During execution, treat changes to the shared agreement as decisions that need acknowledgment. A changed requirement or delivery milestone should prompt negotiation about affected obligations, followed by an updated acceptance condition. Separately measure whether the agents are operational and whether the task is complete. For travel booking, successful messages or tool calls would be intermediate evidence; the requested itinerary would still need verification. A semantic observability layer can help locate disagreements, but its own detection performance should be tested against known failures rather than inferred from a green dashboard.

Pilot selection should consider containment and recovery. Swisscom's reported choice suggests looking for understandable boundaries, redundancy, and rules that can be represented explicitly. Where formal checks are appropriate, ask what they cover, which environmental assumptions they require, and what remains outside the guarantee. Keeping unmodeled cases subject to review reduces the scope of automation, but makes the claimed assurance more defensible.

Access and autonomy also deserve separate decisions. Labenz's laptop agents have broad access with limited autonomy; his Mac Mini agents have narrower access while pursuing longer-running projects. These are two available combinations, not a proven universal prescription. Pandey's description of Arnold adds another dimension: maintaining identity, persona, and memory while changing harnesses and models. A migration can therefore be assessed for continuity without assuming that continuity justifies broader privileges.

Human escalation should have an owner and a purpose. Pandey assigns people judgment and taste under incomplete information, along with accountability for the outcome. A team applying that view could define which conflicts agents may negotiate and which new commitments or unresolved preferences require a responsible person's decision. The dinner example shows the tradeoff: repeated permission requests can consume the benefit of automation, yet removing them does not resolve incompatible preferences. Useful delegation needs both an authorized scope and a boundary at which negotiation stops and escalation begins.

Finally, distinguish available mechanisms from anticipated ones. Pandey separates guardrail-oriented guardian angels from cognitive accelerators, but says the proposed accelerator for reusing past judgment was not being developed at the time. Its future capabilities should not be an acceptance dependency for a current deployment. Likewise, behavioral logs do not establish that latent-space communication is interpretable. Pandey specifically advocates testing and gradual canary deployment for human-facing Neuralink-like connections. Across these cases, the applicable discipline is to evaluate efficiency, observability, and control separately, and keep the deployment boundary within what the available evidence supports.

Source

• Original episode: Agents That Think Together: Cisco's Vijoy Pandey on Semantic Alignment, Scope, Accord & Concord