Cisco chief calls for observability and policy-first approach to scale agentic AI at Across.ai
- Cisco’s AI leadership used an Across.ai appearance to argue that enterprises should scale agentic AI with observability, policy controls, and continuous evaluation. - Cisco’s own AgenticOps push now spans networking, security, and observability, with cross-domain telemetry from Splunk, ThousandEyes, and Secure Firewall. - The message tracks a wider shift from chatbot pilots to governed autonomous systems, as vendors add OpenTelemetry tracing, AI-native logs, and runtime guardrails for agents (cisco.com) (opentelemetry.io) (learn.microsoft.com)
Agentic AI is moving from demos to operations, and Cisco is telling enterprises to treat observability and policy as the control system before they scale. (cisco.com) (learn.microsoft.com) An AI agent is software that can choose steps, call tools, and act toward a goal, not just answer a prompt once. That makes it more useful in IT and business workflows, but also harder to predict and audit. (opentelemetry.io) (learn.microsoft.com) Traditional monitoring watches for crashes, latency, and error rates. Cisco and other infrastructure vendors are arguing that agentic systems also need traces of prompts, tool calls, permissions, retrieval sources, and outcomes, because the same input can produce different actions on different runs. (learn.microsoft.com) (opentelemetry.io) That is the logic behind Cisco’s AgenticOps strategy, which the company introduced in 2025 and expanded on February 10, 2026 across networking, security, and observability. Cisco says the model is built for “autonomous action with built-in oversight.” (cisco.com) Cisco’s case is that scaling agents is not mainly a model problem. It is an operations problem that depends on cross-domain telemetry, governance, and secure execution across cloud, on-premises, industrial, and service-provider environments. (cisco.com) (blogs.cisco.com) The company says its platform draws on live signals from Cisco Networking, Security Cloud Control, Nexus One, Splunk, ThousandEyes, Secure Firewall, and Splunk Observability. Cisco describes that as the context layer that lets agents act with more awareness and less blind automation. (cisco.com) Security is the other half of the pitch. In a February 10, 2026 security update, Cisco said enterprises need controls over what agents can access, which tools they can use, and how to detect manipulation such as poisoned tools or prompts that trigger unauthorized actions. (cisco.com) That policy-first framing lines up with how Microsoft describes AI observability in zero-trust terms: capture agent identity, conversation IDs, tool invocations, permissions, and outputs, then govern retention and access under enterprise policy. In practice, that means role-based access and audit trails become part of the product, not an afterthought. (learn.microsoft.com) The standards layer is still forming. OpenTelemetry said in March 2025 that fragmented agent frameworks were creating a need for shared semantic conventions so telemetry from different tools could be collected in a common format. (opentelemetry.io) Cisco is pushing that same idea in product language rather than standards language: one platform, one view across networking, security, and operations, and one governance model for autonomous systems. Jeetu Patel, Cisco’s president and chief product officer, said in February that AgenticOps combines “cross-domain visibility, purpose-built models, and governance” to run at scale. (blogs.cisco.com) (cisco.com) The immediate takeaway is simple: if companies want agents to do real work, they will need to watch them like production systems and govern them like privileged users. Cisco’s latest product releases show that the vendors selling the pipes, logs, and security controls want to own that layer. (cisco.com 1) (cisco.com 2)
