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

What It Covers

Vijoy Pandey, SVP at Cisco's Outshift division, details how multi-agent systems fail silently in production deployments with partners including Swisscom, ServiceNow, and Webex. He introduces three open-source tools — Scope, Accord, and Concord — designed to detect semantic drift, align agent intent before work begins, and enable negotiation during execution.

Key Questions Answered

• •Silent Failure Rate: Multi-agent systems appear operational while producing zero outcomes. In one travel-booking test with five agents, system dashboards showed 66% uptime — all green — yet task completion was 0%. Cognitive misalignment, not infrastructure failure, caused the breakdown. Operators should not trust uptime metrics alone as a proxy for outcome delivery; semantic observability requires a separate dedicated layer.
• •Cognitive Failure Compounding: Semantic misalignments compound exponentially across agent chains. In a five-agent deployment, one cognitive failure dropped accuracy from 99% to 98%. Two failures dropped it to 81%. Three failures collapsed accuracy to 16% — near random output. Cisco's analysis of public ChatDev traces found 100% of multi-agent sessions contained at least one semantic failure, making this a universal, not edge-case, problem.
• •Scope Protocol — Cognitive Observability: Scope is an open-source observability platform that builds knowledge graphs on top of standard OpenTelemetry traces to detect semantic and cognitive drift between agents. It functions like lane-departure warning software, flagging when agents diverge from shared goals without disrupting existing infrastructure. Teams using any OpenTelemetry-compatible stack can deploy Scope without replacing current monitoring tooling.
• •Accord Protocol — Pre-Work Intent Alignment: Accord is an open-source protocol that forces agents to align on goals, scope, and terminology before execution begins. A concrete failure mode it addresses: one agent interprets "100,000 users" as total users while another assumes concurrent users — a grounding mismatch that cascades through legal, finance, and engineering outputs. Accord rides above A2A and MCP protocols via the open-source Agent Bridge project.
• •Concord Protocol — Mid-Work Negotiation: Concord handles coordination and compromise during active task execution, where conditions shift and agent assumptions diverge. In the restaurant-booking example, two agents representing different people with conflicting food preferences, location requirements, and availability windows cannot autonomously reach a reservation without Concord-style negotiation. The protocol enables agents to update shared understanding dynamically rather than locking in initial assumptions.
• •Identity-Model Decoupling Strategy: Pandey separates agent identity, persona, and memory from both the underlying model and the harness. His personal agent "Arnold Lane" runs simultaneously on Hermes, Muse, DOTS, and Grok — switching models without losing continuity. This architecture prevents vendor lock-in and survives model deprecations. The practical implementation: store persona and memory in a layer that is explicitly independent of whichever frontier model or harness sits beneath it.
• •Latent Space Communication — Inevitable for Hard Science: Agent-to-agent communication in high-dimensional latent space, bypassing natural language entirely, is positioned as necessary for problems like drug discovery, materials science, and weather modeling where token-based communication introduces lossy compression. Pandey frames this as analogous to pharmaceutical drug introduction — the capability should be developed and deployed incrementally with staged rollouts, not blocked, but not rushed into human-facing systems without structured guardrails.