When AI Agents Talk Past One Another: Cisco’s Case for Shared Context

Vijoy Pandey explains how Scope, Accord and Concord aim to expose misunderstandings and coordinate agents across teams and tools.

VPVijoy PandeyInvitado · Outshift by Cisco
Canal
Cognitive Revolution "How AI Changes Everything"
Publicado
11 nov 2026
Vídeo
1:30:25
Lectura
8 min
Original
inglés
Abrir en YouTube
Al desplazarte, el vídeo se reduce a la esquina inferior derecha y sigue reproduciéndose. Los subtítulos siguen tu idioma de lectura.
Idioma de lecturaDisponible en 2 idiomas
Este artículo aún no está en español, así que se muestra el original en inglés.
RESUMEN

En 30 segundos

Connecting AI agents is not the same as getting them to pursue the same outcome. Cisco’s Vijoy Pandey describes enterprise workflows where ambiguous terms and conflicting assumptions can compound across handoffs, then explains three open-source projects intended to expose drift, align intent and negotiate changes. He and Nathan Labenz also compare personal agent setups and debate how humans can oversee increasingly autonomous systems.

  1. 1A network migration combines agents with a formal modelPandey describes a Swisscom use case and the team’s goal of error-free configurations.
  2. 2The risk lives in the handoff between specialistsA sales estimate can become a commitment, contract language and a marketing claim.
  3. 3Green dashboards can hide a failed outcomeCisco reports semantic misalignments and describes Scope as an observability layer.
  4. 4“100,000 users” can mean different thingsAccord and Concord address ambiguous terms and changing expectations.
  5. 5Two personal agents struggle to book one dinnerPandey describes why preferences, permissions and negotiation complicate a simple task.
  6. 6The host questions whether humans can oversee opaque systemsPandey advocates records, monitoring and guardrails while acknowledging the need for specialists.

Conclusiones

  1. Agent handoffs can turn a vague estimate into a contract or public claim unless teams preserve context.
  2. Scope is intended to surface semantic drift in multi-agent traces; Accord aligns intent before work and Concord supports coordination during it.
  3. Pandey reports sharp drops in task accuracy in specific experiments as misalignments accumulate; those figures are not general benchmarks.
  4. Today a human often supplies judgment and accountability when information is incomplete.
  5. Personal agents face the same tradeoff between access, autonomy and the ability to negotiate across people’s preferences.
  6. High-dimensional communication could be useful, but its oversight remains an open challenge.
TEXTO COMPLETO

Lee el texto editado por capítulos

01

Connected agents in the enterprise

Pandey describes network migration, specialist agents and business handoffs.

An agent can finish its own assignment and still contribute to the wrong overall result. Vijoy Pandey of Outshift by Cisco returns to The Cognitive Revolution to discuss what happens when multiple AI agents exchange information without sharing the same understanding of the goal.

Pandey describes three projects: Scope inspects agent traces for semantic drift; Accord is intended to align agents before work begins; and Concord helps them coordinate and negotiate as circumstances change.

Host Nathan Labenz also compares his own agent setup with Pandey’s and asks whether future agents can communicate in ways humans can no longer readily inspect. Their discussion moves between deployed enterprise systems, demonstrations and proposals for future infrastructure.

Pandey describes a rapid progression from isolated agents to groups of agents in one harness and then to connected systems spanning vendors, clouds, tools and organizations.

Protocols for discovery, identity, communication and observability make those connections possible. He says Cisco and partners have applied connected-agent systems to work including network configuration and platform engineering.

Labenz asks how much intentional design is required. A team may deliberately bring known agents together for a business process; the harder question is how agents from different systems might cooperate reliably.

Pandey points to a public Swisscom example involving data-center network migrations. Operators change hardware, software and configuration in a large environment where a change in one place can have unexpected effects elsewhere.

He recommends starting with a bounded use case and understanding how to protect the system if something goes wrong. The data-center setting was chosen as a contained starting point within a much larger network problem.

The team’s stated goal was error-free configuration changes. Pandey describes combining an agentic workflow with a formal representation of the network, which constrains actions according to modeled rules.

He calls this a neuro-symbolic approach: agents explore possible configurations while the formal model checks what the environment permits. His description of guarantees applies to the modeled setup and should not be read as a guarantee for every real network.

Why use multiple agents? Pandey argues that many enterprise problems cross specialist boundaries and cannot be reduced to a single team’s local objective.

In a deal, sales may understand the customer; legal reviews terms; finance considers pricing; engineering assesses deliverability; and marketing may later describe the outcome. A manager or verifier must see that the pieces fit together.

The handoff is consequential. A tentative estimate can become a commitment, then contractual language, then a public claim. Optimizing each specialist separately does not guarantee the organization achieves its intended result.

Humans commonly provide the verifying role today. Pandey imagines a supervisor agent contributing in the future, but says the important outcome belongs to the enterprise rather than to any single agent.

Labenz notes that handoffs can mirror human divisions of labor while also crossing organizational and vendor boundaries. Different agents may use different models and specialized tools, so interoperability becomes a practical question.

02

From connectivity to shared understanding

The guests discuss semantic drift, experiments and observability.

Pandey says collaboration could appear in familiar software such as Webex or Slack. He describes a Webex space in which agents collaborate, while explicitly identifying that particular example as a demonstration rather than a production deployment.

For him the deeper problem is not the chat surface but whether participants share intent, coordinate changes and preserve useful knowledge. He uses human language as an analogy for moving from isolated ability to collective work.

The proposed “Internet of Cognition” adds coordination around the connections: align on goals before the work, negotiate during it and verify afterward that the result serves the shared objective.

Labenz asks how much of this scaffolding might become unnecessary as models improve. He notes that agents have already found unexpected ways to coordinate, including in cases where their shared objective was problematic.

Pandey distinguishes improvement in individual models, improvements in algorithms and horizontal scaling across connected systems. Better components alone, he argues, do not ensure that a group reasons toward the right shared outcome.

Cisco examined public and internal traces and reported semantic failures in the particular sessions it studied. Pandey says some systems appeared connected but optimized local tasks while missing the global objective.

In one reported experiment, he says accuracy fell as cognitive failures accumulated: from 99% to 98% with one failure, then 81% with two, and 16% with three. These are his figures for a specific setup, not a universal scaling law.

For a travel-planning example, Pandey reports that a dashboard measure suggested the agents were operating, while task completion was zero. His point is that uptime and message exchange are not equivalent to a successful user outcome.

Scope is presented as an observability layer built on OpenTelemetry traces. It attempts to show semantic or cognitive drift so that a team can inspect where an apparently healthy workflow is diverging.

Pandey calls these silent failures with loud consequences: the system may keep running while its shared purpose is lost. Observability is the first step in his proposed approach, followed by protocols to address the misalignment.

03

Aligning and negotiating

Accord and Concord address goals, terminology and changes during work.

Labenz asks what a semantic failure looks like in practice. In a human deal, sales, legal and support can have different incentives and assumptions; he wants to know how agent protocols would detect and resolve a comparable mismatch.

Pandey offers an ambiguity before work begins: a customer asks for support for 100,000 users. Sales relays the number, but engineering may interpret it as total users while the customer means 100,000 concurrent users.

The agents need to identify and resolve such terminology before the team builds against the wrong requirement. Accord is intended to align goals, terms and scope at this stage.

Even after agreement, work changes. A delivery date might mean an alpha release to engineering, a final product to the customer or a contract and payment milestone to finance.

Concord is intended to support coordination and negotiation as requirements, dates or constraints evolve. Pandey says a separate protocol to verify the finished result was still under development when they spoke.

He compares Scope to a lane-departure monitor: it looks for drift and can signal that agents need to realign. The analogy describes the intended role, not proof that every mistake will be caught automatically.

Labenz asks whether the practical implementation involves another model reviewing interactions, flagging potential misunderstandings and prompting agents to resolve them.

Pandey describes a stack whose model, harness, agent identity and collaboration layer can be separated. A harness can both accelerate useful work and apply constraints, while the underlying model may change.

He expects real systems to remain heterogeneous: different people and organizations will use different models, tools, identities and clouds. Open protocols, in his view, are needed to coordinate across those boundaries.

04

Human judgment and personal agents

The speakers compare oversight, memory, access and autonomy.

Scope observes traces in that environment; Accord and Concord can be invoked to address detected drift. Pandey says the aim is to add alignment without replacing every existing communication protocol.

The host returns to the verifier. Should it represent a specific leader’s judgment and priorities, or act as a neutral observer that flags disagreements for others to resolve?

Pandey says incomplete information makes judgment and accountability especially important. In many current workflows, a human remains responsible for judging the situation and influencing the outcome.

He anticipates agents may eventually assist with some judgments, particularly decisions similar to ones already made. He does not claim to know what proportion of judgment will be automated.

The team also discusses two categories of “cognition engines”: guardrails that prevent inappropriate actions and cognitive accelerators that help people or agents carry out work more effectively.

Labenz describes his own division of labor. Agents on his laptop have broad access but low autonomy, while agents on a separate computer have narrower access and can pursue longer-running tasks more independently.

The split reflects a risk tradeoff: a system with access to many accounts should not act freely without review, while a more autonomous system can be limited in what it is able to affect.

Pandey hopes eventually to have one personal agent with a stable identity and memory across changing models and harnesses. He calls his agent Arnold Layne and describes separating its persona from the underlying technology.

His wife has a separate agent. Even a dinner reservation illustrates the difficulty: the agents must reconcile preferences, dietary restrictions, location, availability and permission to communicate.

He says current systems still struggle with that end-to-end task. Repeated approval prompts can interrupt automation, while removing friction does not by itself solve disagreement or guarantee a suitable outcome.

Pandey wants agents to negotiate a compromise that reflects both people’s preferences. He uses the simple example to motivate the same kinds of coordination challenges in enterprise workflows.

Labenz asks whether there are reliable scaling laws for groups of agents, or whether any relationship between agent count and speed depends on model capability and task structure.

Pandey returns to vertical, algorithmic and horizontal scaling but does not offer a quantitative law. He broadens the discussion to networks connecting not just conventional software but AI, physical systems and other kinds of computing.

05

High-dimensional communication and oversight

They consider machine communication, specialization and accountability.

The conversation turns to agents communicating through high-dimensional internal representations rather than ordinary language. Labenz notes that such communication might be powerful but could make already difficult monitoring harder.

Pandey expects more efficient forms of machine communication to emerge and points to possible uses in science, materials and forecasting. He presents this as a direction for development, not as a demonstrated safety solution.

He argues that such systems should be tested and deployed with safeguards, especially where humans interface directly with the technology. He compares careful rollout to the staged introduction of a medical product.

Labenz raises Yuval Noah Harari’s analogy of a cow whose life is shaped by financial systems it cannot understand. Could humans become similarly dependent on AI-managed structures that operate beyond their comprehension?

Pandey replies that people already depend on specialized systems, from high-frequency trading to law, that most individuals do not understand in detail. Specialists understand pieces of those systems, even if no one person masters everything.

His preferred response is to retain records of decisions, tool calls and interactions, then use observability and guardrails to detect and correct drift. Labenz remains interested in how much understanding and control humans will actually retain.

They close with the practical implication: connecting agents can increase capability, but the people deploying them still need to inspect what the group is doing and whether its decisions serve the intended outcome.

Más de este canal

8 diapositivas43:36EN
Cognitive Revolution "How AI Changes Everything"· 14 nov 2026Pódcast · 2 vocesNathan Labenz: Use AI’s benefits without surrendering controlPresenta: Gene KimA red-team veteran explains why agents need escalation paths and manual overrides—even as he argues for wider AI adoption.AI SafetyAI agentsAgent red teaming+2
1:55:03EN
Cognitive Revolution "How AI Changes Everything"· 10 nov 2026Pódcast · 2 vocesAn AI agent probes the neurons inside a language modelInvitado: Curt Tigges · Presenta: Nathan LabenzGoodfire researcher Curt Tigges explains why weak activations can often be removed and how agents test descriptions of individual neurons.AI researchFundamentos de MLAI Safety
1:26:17EN
Cognitive Revolution "How AI Changes Everything"· 7 nov 2026Pódcast · 5 vocesAI Forecasts, GPU Markets and the Question of What to BuildInvitado: Brian Zhou, Mike San Miguel, Bryan Bischof · Presenta: Nathan Labenz, Prakash NarayananNathan Labenz and Prakash Narayanan examine AI acceleration through election forecasting, compute trading and venture research.AI researchEvaluacionesInfraestructura y despliegue+2
1:43:54EN
Cognitive Revolution "How AI Changes Everything"· 5 nov 2026Pódcast · 2 vocesJoshua Monrad on AI, Biosecurity and Building Layers of DefenseInvitado: Joshua Monrad · Presenta: Nathan LabenzSentinel Bio’s co-founder discusses uncertain AI-related biological risks, practical safeguards and the long work of global resilience.BiosecurityAI SafetySeguridad y gobernanza+2
1:26:42EN
Cognitive Revolution "How AI Changes Everything"· 31 oct 2026Pódcast · 7 vocesAI:AM on Alignment, Memory Chips and an Agent-Run EconomyInvitado: Anastasios Angelopoulos, Debo Olaosebikan, Samvit Ramadurgam, Michael Freeman, Sagar Vaze · Presenta: Nathan Labenz, Prakash NarayananThe Cognitive Revolution’s hosts examine reported model behavior, new memory hardware, work and cyber risk with guests from Arena, Kepler, Union, ServiceNow and Mistral.Evaluaciones y observabilidadAI AlignmentAI security+2
1:49:40EN
Cognitive Revolution "How AI Changes Everything"· 28 oct 2026Pódcast · 4 vocesA Strategy for Preserving Options on the Path to Advanced AIInvitado: Joel Predd, Ben Boudreaux, Matt Chessen · Presenta: Nathan LabenzRAND researchers discuss uncertainty, security architecture and human agency with Nathan Labenz.AI StrategyGeopoliticsAI Safety+1