In today’s hyper‑connected enterprise, the pressure on network operations teams has never been greater. Imagine a Formula 1 pit lane during a Grand Prix: engineers stare at live telemetry, strategists run simulations, mechanics prepare for sub‑second tire changes, and data analysts flag anomalies before they threaten the car’s performance. Success hinges on every specialist acting in concert, reacting to shifting conditions in real time. Modern networks mirror that scenario—traffic spikes, AI‑driven workloads, and a patchwork of vendor equipment create a constantly shifting landscape where delays or missteps cascade into service degradation. The old playbook of automating isolated tasks falls short because it treats the network as a collection of independent levers rather than a tightly coupled system. To keep pace, organizations need a framework that enables intelligent, coordinated decision‑making at machine speed, turning a disparate set of tools into a cohesive, self‑optimizing crew.
Many network teams today rely on scripts, point solutions, and rudimentary orchestration that address only one symptom at a time. When a latency spike appears, an engineer may manually pull logs from routers, switches, and firewalls, then correlate timestamps across disparate dashboards. This reactive hunt consumes precious minutes, often after users have already felt the impact. Meanwhile, telemetry volumes explode as streaming sensors, gNMI feeds, and BGP‑LS updates flood the environment, while operational data remains trapped in silos—each domain storing its own metrics, events, and configuration states. The result is a bottleneck where human analysts become the limiting factor, unable to synthesize the firehose of information fast enough to prevent outages. Even when automation exists, it tends to be brittle, breaking when a new device type or a sudden traffic pattern deviates from pre‑programmed rules. The core issue is a lack of true teamwork among automated components; they act as soloists rather than an orchestra guided by a shared score.
Cisco’s answer to this challenge is the next evolution of its Crosswork Network Automation suite: Crosswork® AI, a secure, scalable multi‑agentic framework designed to operate as an extension of the network team itself. Rather than delivering another monolithic tool, Crosswork AI introduces a collection of specialized software agents, each endowed with a distinct function—much like the engineers, strategists, pit crew, and analysts in a racing team. These agents continuously ingest telemetry, exchange insights, and jointly reason about the network’s state. By embedding the framework within the existing Crosswork automation backbone, organizations gain a platform that can scale across thousands of devices while preserving low‑latency communication between agents. The result is a shift from isolated scripts to a living, learning team that can anticipate problems, validate intentions, and execute corrective actions before users notice a hiccup.
The power of the multi‑agentic approach lies in how the agents collaborate. One agent might focus on real‑time anomaly detection, streaming statistical models that flag deviations from baseline behavior. Another could specialize in intent validation, constantly checking whether the configured network state matches the business objectives expressed by orchestration platforms. A third agent may act as a troubleshooting expert, traversing the knowledge graph to hypothesize root causes and suggest mitigation steps. Meanwhile, a planning agent simulates the impact of proposed changes, using predictive analytics to forecast congestion or security exposure. Because each agent communicates through a shared message bus and references a common knowledge graph, insights flow rapidly: the detection agent alerts the troubleshooting agent, which consults the planning agent to evaluate safe remediation paths, and the intent agent confirms that the proposed fix does not violate service‑level agreements. This tight loop enables decision‑making at speeds that match the network’s own dynamics.
To make this collaborative intelligence visible and actionable, Crosswork AI is tightly integrated with Cisco AI Canvas, the generative user interface that forms part of Cisco Cloud Control. AI Canvas transforms raw agent outputs into intuitive visualizations, natural‑language summaries, and interactive dashboards that network engineers can explore without deep data‑science expertise. Imagine a screen where a simple query like “Show me any potential SLA breaches in the next 15 minutes” triggers a cascade: the intent‑validation agent scans upcoming schedule changes, the anomaly‑detection agent scans live telemetry, and the planning agent runs what‑if scenarios—all within seconds. The result appears as a concise briefing, complete with recommended actions and confidence scores. By embedding the multi‑agentic logic inside a generative UI, Cisco lowers the barrier for teams to adopt AI‑driven insights, turning complex probabilistic reasoning into clear, operational guidance that can be acted upon immediately.
Recognizing that no two networks are identical, Cisco built Crosswork AI to be extensible from the ground up. The framework ships with an Agent Development Kit (ADK) that lets partners, customers, or internal teams create custom agents tailored to proprietary protocols, niche monitoring tools, or specialized business logic. Once developed, these agents can be registered in the internal Agent Catalog, where they undergo automated evaluation against performance, security, and compliance criteria. A central Knowledge Graph stores relationships between devices, services, policies, and historical incidents, giving every agent a shared contextual foundation. This architecture means that a team can start with Cisco’s out‑of‑the‑box agents for common functions—such as fault detection, capacity planning, or security posture—then gradually layer in bespoke agents that address unique challenges, like a legacy signaling protocol or a custom application‑level health check. The plug‑and‑play nature ensures that the framework evolves alongside the network, protecting investments while encouraging innovation.
The industry has already taken notice of this approach. At a recent TM Forum Catalyst showcase, the Crosswork AI multi‑agentic framework was featured in the “Game X” project, which was selected as a Moonshot—an accolade reserved for initiatives that promise transformative impact on the telecommunications sector. Being named a Moonshot signals that experts view the technology not as an incremental improvement but as a potential catalyst for redefining how service providers design, operate, and monetize their networks. The recognition highlights several key strengths: the framework’s ability to handle massive scale, its reliance on open standards for interoperability, and its focus on closing the loop between intent, automation, and assurance. For enterprises and service providers alike, such validation reduces perceived risk when considering adoption, offering an external endorsement that the multi‑agentic model can deliver measurable improvements in mean time to resolve, operational expenditure, and service agility.
Beyond isolated use cases, the true value of Crosswork AI emerges when it is applied across the entire network lifecycle. From the moment a new device is unboxed, the onboarding agent can verify firmware compatibility, push baseline configurations, and register the asset in the inventory database. As services are designed, the planning agent simulates traffic patterns, validates capacity, and checks compliance with security policies. During rollout, the orchestration agent coordinates provisioning steps across multi‑vendor equipment, while the assurance agent continuously monitors KPIs and raises alerts if deviations appear. Optimization agents work in the background, tuning load‑balancing parameters, adjusting QoS policies, and recommending energy‑saving measures. Finally, when an incident occurs, the troubleshooting agent leverages the knowledge graph to correlate events, isolate the fault domain, and suggest remediation steps—often before the issue escalates to a customer‑impacting outage. This end‑to‑end coverage ensures that intelligence is not confined to a single phase but permeates every operational touchpoint.
A foundational capability of the framework is its ability to translate business and operational intent into concrete network state, continuously and autonomously. Intent‑driven networking has long promised that operators could declare what they want—such as “Provide low‑latency connectivity for all video‑conferencing traffic between headquarters and regional offices”—and let the system figure out how to achieve it. Crosswork AI makes this promise operational by maintaining a live intent repository that is constantly reconciled against the actual configuration and telemetry data. When a change request arrives, the intent agent validates it against business rules; the planning agent models the impact; and the orchestration agent executes the necessary adjustments, while the assurance agent monitors post‑change performance. If drift occurs—perhaps due to a manual patch or a hardware replacement—the detection agent flags the discrepancy, and the remediation loop initiates corrective actions without human intervention. This closed‑loop behavior ensures that the network remains aligned with strategic goals even as conditions evolve.
Modern networks rarely consist of equipment from a single vendor, making open, standards‑based interoperability a prerequisite for any automation platform. Crosswork AI leans heavily on industry‑recognized interfaces such as NETCONF, RESTCONF, TAPI, OpenConfig, and native YANG data models to issue commands and collect state. For real‑time visibility, it subscribes to streaming telemetry protocols like gRPC, BGP‑LS, and gNMI, which deliver sub‑second updates on interface counters, routing tables, and resource utilization. By adhering to these open contracts, the framework avoids locking customers into a proprietary stack, allowing them to mix and match best‑of‑breed gear from multiple suppliers. This openness also simplifies upgrades: when a new device model ships with updated YANG modules, the existing agents can typically consume the new data without requiring rewrites, preserving operational continuity while enabling technology refresh.
Real‑world validation reinforces the claim that Crosswork AI can handle heterogeneous environments, including legacy and proprietary systems that do not natively support modern APIs. In the European Advanced Networking Test Center (EANTC) interoperability showcase, the framework successfully automated end‑to‑end services across a live, mixed‑vendor topology that combined current‑generation routers, older SDH equipment, and specialized optical transport gear. For components lacking standard interfaces, Crosswork provides an open SDK that enables partners to develop custom packages and agents capable of speaking the device’s native CLI, SNMP, or proprietary SOAP‑based APIs. These adapters plug into the same message bus and knowledge graph, ensuring that even the most idiosyncratic gear participates in the multi‑agentic decision‑making process. Consequently, organizations can protect their existing investments while gradually modernizing, knowing that the automation layer will not leave any domain behind.
Looking ahead, Cisco plans to showcase the full capabilities of Crosswork AI at Cisco Live 2026 in Las Vegas, inviting attendees to experience the multi‑agentic framework in action within the Secure Global Connectivity area of the World of Solutions. For network leaders eager to begin the journey, a practical first step is to run a pilot focused on a high‑value use case—such as automated anomaly detection for critical business services—and measure key performance indicators like mean time to detect, mean time to resolve, and operational effort saved. Simultaneously, invest in upskilling teams on intent‑based networking concepts and the basics of agent development, so they can contribute custom agents that address unique organizational needs. By treating the network as a coordinated team of intelligent agents, rather than a collection of isolated tools, enterprises can move from reactive firefighting to proactive, continuous optimization—operating with the precision, intelligence, and coordination of a championship‑caliber pit crew.