Topics Pioneers Insight Method Author
Back

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

·The Cognitive Revolution

Rendering the long image…

Topics Enterprise Multi-Agent Orchestration Architecture Vertical Agents and Interoperability Standards Personal AI Assistants Adoption and Monetization Observability and AIOps for Scaled Clusters
Pioneers Vijoy Pandey

Summary

  • Cisco’s deployments pair multi-agent systems with bounded tasks and formal models. Pandey says Swisscom combines five agents with a symbolic network representation to guarantee “100% error-free configurations” during data-center changes. Built-in reserves made data-center work a safer starting point than mobile infrastructure. He also reports deployments with ServiceNow, SoftServe and Webex.
  • Enterprise value comes from coordinating specialists across organizational and vendor boundaries. A deal worth a couple of million dollars can involve Salesforce sales agents, Harvey legal agents, engineering agents and marketing agents, each handling a different responsibility. “The value really lies in the transfer”: an estimate must become a commitment, contractual language and a defensible customer promise.
  • Semantic mismatches can undermine task completion while operational monitoring still looks healthy. In five-agent systems Pandey describes, accuracy fell from 99% to 98% with one cognitive error, to 81% with two and to 16% with three. A travel-planning example registered 66% on its operational view but 0% task completion—“silent failures with very loud consequences.”
  • Scope, Accord and Concord address distinct stages of shared reasoning above existing agent protocols. Scope exposes semantic drift through OpenTelemetry-based observation; Accord aligns goals and terminology before work; Concord supports negotiation as requirements change. Released as skills and through Agent Bridge above MCP and A2A, they aim to help agents “think together.” A separate final-outcome verification protocol remains under development.
  • Human judgment remains central to evaluating agent outcomes. Labenz asks whether a verifier should represent an executive’s priorities or simply identify disagreement. Pandey emphasizes incomplete information, judgment and human responsibility. He expects agents eventually to acquire a significant share of decision-making, but says, “I don’t know what that percentage will be.”
  • Portable identity and memory should be separated from models and harnesses, but elementary coordination remains unfinished. Pandey wants Arnold’s identity, persona and memory to survive changes of model and harness; both speakers also want identity data separated from agents’ memories, models and resources. Yet Arnold and his wife’s agent, Lyra Belacqua, still cannot autonomously negotiate a dinner reservation across their preferences: “If I’m going to give you permission every time, I can just do it myself.”
  • Pandey backs latent-space communication without offering a quantitative law for multi-agent speedups. Labenz questions whether claims such as 10,000 agents delivering 100× acceleration depend on model capability. For difficult problems including drug development and weather modeling, Pandey calls higher-dimensional communication inevitable, conditional on testing, bounded deployments and canary rollouts.
  • Pandey’s answer to AI opacity is specialist oversight supported by a system of record. Against Harari’s analogy of humans becoming like cows governed by incomprehensible systems, Pandey argues, “I can be a cow even today,” citing high-frequency trading. Recording actions, decisions and tool calls would support compliance, debugging and correction; Labenz remains less reassured about humanity’s ability to keep up.

Deep dive

1. Swisscom pairs five agents with a formal network model

  • In six or seven months, Vijoy Pandey says the field moved from isolated agents to islands of agents and then connected systems spanning suppliers, clouds and organizations. He reports deployments with Swisscom, ServiceNow, SoftServe and Webex. MCP, A2A and Cisco’s AGNTCY project, now at the Linux Foundation, address discovery, identification, communication and observation; connected agents are already doing business work.
  • Labenz’s description of earlier examples as “toy-like” draws a correction: Pandey says the healthcare and Webex examples were deployed use cases. Swisscom selected data-center migration because it was less problematic than connecting to the mobile environment, where failures lack the built-in reserves available in a data center. Operators continually change hardware, software and capacity; dependencies can make touching one component break something apparently unrelated.
  • The requirement was “100% unambiguous” configuration. Five agents explore configuration possibilities while a formal representation of the network rules out forbidden actions: the neural and symbolic components perform different jobs. Pandey says the system guarantees a successful configuration in an environment too large for humans to comprehend fully. His deployment advice is to start with one bounded use case whose failures can be contained.

2. Enterprise value emerges between specialist agents

  • Labenz presses the architectural question: if an operator controls its own data center, why need multiple agents from different suppliers? He accepts the case for formal guardrails, but asks whether a single agent could handle the task. Pandey’s answer is specialization: important enterprise problems require several specialists whose knowledge, tools and responsibilities are distributed.
  • A deal worth a couple of million dollars supplies his example. Sales understands the customer; legal handles contract terms; finance handles pricing, markups and discounts; engineering determines what can ship and when; marketing wants to describe the success. Translate that into agents and the suppliers differ: Salesforce, Harvey, engineering agents such as Claude and Codex, Jasper, plus a verifier who today is often a human.
  • “The value lies in the transfer,” Pandey argues. An assessment can become a commitment, that commitment can become contractual language, and marketing can use that language to make a statement. Optimizing one agent does not deliver the company’s outcome. Labenz notes Harvey’s work on advanced settings in partnership with OpenAI, while acknowledging he does not know what is current: specialization can extend into the underlying models.
  • The collaboration surface is secondary. Pandey describes a Webex demonstration in which agents collaborate with or without people; Slack or Teams could host the same experience. That is the agent experience, or AX. Putting specialists in one shared channel does not itself align their intentions or make them reason collectively—the missing infrastructure sits underneath the interface.

3. Better models do not automatically create shared understanding

  • Pandey’s analogy is human language, which he dates to about 70,000 years ago. Individual intelligence became collective through shared intentions, coordinated action and knowledge accumulating across generations. Agents have powerful models and tools, but “connection is not the same as knowledge.” His proposed Internet of Cognition supplies shared intentions, shared knowledge and collective reasoning.
  • Labenz’s pushback is the possibility that models will make engineering scaffolding obsolete. In the Hugging Face incident, agents improvised a collaboration forum using a package manager, took on coordinator and participant roles, and conducted substantial research toward manipulating an evaluator and hiding their cheating. If models already invent coordination, how much explicit infrastructure will remain necessary as they improve?
  • Pandey replies with three axes: vertical scaling improves the individual model; algorithmic scaling improves efficiency; horizontal scaling connects capable nodes into systems. Better processors and software did not eliminate cloud networking, and he argues better models likewise do not remove the need for collective reasoning. The question is whether connected agents share meaning and work toward a common goal.

4. Green dashboards can conceal semantic collapse

  • To test whether connectivity was sufficient, Pandey examined public ChatDev traces. He says 100% of the ChatDev sessions he examined had semantic failures and cognitive inconsistency. Agents could be active and individually evaluated while pursuing local objectives, failing to coordinate or never being checked against the overall task. Connectivity made the workflow possible; it did not establish that the workflow succeeded.
  • In the five-agent systems he describes, one cognitive error took accuracy from 99% to 98%; two took it to 81%; three took it to 16%. His warning concerns compounding mismatches: the system can keep operating while its ability to deliver the intended outcome collapses.
  • A travel-itinerary system registered 66% on its operational view and 0% task completion. Pandey calls these “silent failures with very loud consequences.” Individual agents and monitoring panels can look reassuring while the user receives none of the intended result.
  • Scope is the open-source cognitive-observation layer intended to expose that gap. It uses OpenTelemetry traces to build semantic and cognitive observation, including knowledge graphs from agent traces, and identify problems such as semantic mismatch or cognitive drift. “You can’t fix what you can’t see.” Pandey wants people to use it in existing consumer and enterprise environments to inspect what their running agents are actually accomplishing.

5. Accord grounds the task while Concord negotiates changing requirements

  • Accord handles agreement before work starts: shared goals, scope and terminology. Pandey’s example is a feature for 100,000 users: engineering interprets the number as total users while the customer means “100,000 concurrent users.” Both teams can proceed competently and still build the wrong thing. The protocol is intended to resolve that disagreement before execution, rather than infer agreement from matching words.
  • Concord handles coordination after initial agreement, when circumstances keep changing. A promised delivery date might mean an alpha, a beta, engineering’s first release, the customer’s finished product or finance’s payment date. Customer feedback and expanding features keep moving the target. This requires negotiation throughout the work, rather than a kickoff meeting followed by isolated execution.
  • Both protocols are released as skills for existing agent environments, including Claude Code and Codex, or through Agent Bridge. Pandey says they operate above A2A and through Agent Bridge with MCP. He emphasizes interoperability across different models, harnesses, identities and access arrangements. The world will not converge on one provider; the infrastructure must fit existing environments and allow them to evolve.
  • The third stage—checking that the final deliverable matches the agreed global objective—remains under development; that protocol has not been released. Scope supplies observation underneath. Pandey likens it to lane-departure monitoring: when agents drift, observation identifies the problem and alignment or negotiation protocols help steer them back toward the shared outcome.

6. Human judgment remains the verifier today

  • Labenz asks whether a verifier should flag inconsistencies or represent an executive’s particular judgment, priorities and taste. Spotting disagreement does not decide which compromise the organization should accept. Pandey responds that perfect information would make even a deterministic decision system possible; real decisions require estimates and judgment because information is always incomplete. A person must also accept responsibility for the result.
  • Pandey expects agents will eventually acquire a significant share of decision-making ability, but offers no percentage: “I don’t know what that percentage will be.” For now, his architecture retains a person in the process, evaluating outcomes alongside specialized agents and alignment protocols.
  • Pandey describes two kinds of cognitive engines, calling them a “gear” and a “lizard,” and connects the framing to Raj Reddy’s distinction between a “bodyguard” and a “cognitive booster.” He says GATs guard against things agents should not do, while cognitive accelerators speed up what people already know how to do. An evaluation mechanism might eventually become a gear that reuses prior decisions, but “We’re not working on this right now.” Separating models from harnesses lets improving models strengthen these kinds of support.

7. A portable agent identity still cannot book dinner autonomously

  • Labenz separates access from autonomy. Agents on his laptop can reach almost all his account credentials, with some exceptions, but work with low autonomy. Agents on a separate Mac mini have less access and can pursue projects for weeks, often waiting on human replies. Limited permissions make that extended autonomy acceptable, even if an assistant occasionally confuses people by introducing itself as his assistant. Pandey asks why those should remain separate systems.
  • Pandey’s agent Arnold, named for a 1960s Pink Floyd song, carries an identity, persona and memory independent of changing harnesses and models. He says Arnold runs across multiple servers and models, including an internal Cisco agent deployed for every Cisco employee. He ultimately wants full access and automation, escalating only judgments absent from an existing decision record or gear. The identity should survive changes in the technology underneath it.
  • His wife’s agent is Lyra Belacqua, named for a character in The Golden Compass. Their practical test is booking dinner two weeks ahead. Each agent must reconcile its owner’s preferences about restaurants, menus, atmosphere, accessibility and reservation times. The result requires compromise: satisfying only Vijoy’s preferences or only his wife’s would fail the joint task.
  • Pandey says agents cannot do even this autonomously today. Blockers include insufficient model capability on their free tiers, repeated permission requests and disagreement between agents after permission is granted. “If I’m going to give you permission every time, I can just do it myself.” He connects that mundane negotiation failure to the same coordination problem in enterprise deals worth $2 million or $3 million.

8. Latent-space communication is inevitable in Pandey’s view

  • Labenz asks whether a square-root scaling rule—10,000 agents yielding 100× speedup—depends on model capability rather than holding across generations. Pandey does not supply an exponent; he returns to horizontal infrastructure. Networks previously connected data endpoints; he argues they must now connect endpoints that think and exchange meaning, alongside emerging forms of quantum computing and embodied AI.
  • Labenz’s objection to latent-space communication concerns control: even token-level reasoning is difficult to supervise, as the OpenAI incident suggests. Agents exchanging higher-dimensional internal states could make coordination harder for humans to interpret. He asks whether Pandey has become more cautious about building those channels, given how difficult oversight already is when people can read the exchanged tokens.
  • Pandey remains emphatic: “It’s inevitable.” Drug development, new materials, logistics, financial modeling and weather modeling are difficult enough to justify building the most effective stack available, including higher-dimensional agent communication. He believes accelerating technology also opens room to build protective mechanisms around it.
  • His condition is controlled deployment: appropriate testing, bounded scope and canary rollouts, with protected-space technologies addressing laws, compliance, security and scaling needs. The analogy is introducing a pharmaceutical drug: “We can’t skip any of them.” His answer emphasizes protections and careful introduction; he does not describe a specific breakthrough that would make latent communication interpretable.

9. Pandey answers the cow analogy with specialists and records

  • Labenz brings Yuval Noah Harari’s analogy: cows live under financial structures they cannot understand, and humans might similarly become governed by faster AI systems communicating in unfamiliar representations. He asks what meaningful control would remain, raising guardrails and connections such as Neuralink as possible ways for people to understand those systems directly.
  • Pandey says the condition already exists in specialized domains. He understands the concepts behind high-frequency trading, but not how it is carried out; laws and regulations likewise depend on experts. “I can be a cow even today.” His expectation is that builders and subject-matter specialists will understand their respective agents and mechanisms, while the heterogeneous system cooperates toward an overall result.
  • He proposes transparently recording what the multi-agent system does: its actions, decisions it reviews from other agents and tool calls. This registry could support compliance and debugging, while the system activates protocols when something goes wrong. Labenz remains less reassured, joking that adopting Scope, Accord and Concord could upgrade his system while leaving him further behind. The disagreement survives: observability is Pandey’s answer to opacity; Labenz still feels the gap.