Uploaded October 2026 | Updated November 2026, 1 day ago
AI agents are rapidly moving beyond chat-based assistants toward systems that can reason over data, interact with tools, and automate operational workflows. But applying these capabilities to network troubleshooting introduces unique challenges: network state is distributed across multiple systems, telemetry may be incomplete or stale, and an incorrect action can have significant operational impact.
This talk walks through the design and implementation of an AI agent for network troubleshooting, using a virtual router environment as a practical example. We will explore how to define the right problem for an agent, design its architecture, expose network capabilities through tools, and enable the agent to investigate routing and connectivity issues using real network state.
The talk will also focus on the engineering considerations that become critical when deploying agents in network operations, including tool selection, context management, deterministic workflows versus LLM-driven reasoning, guardrails, permissions, human approval, failure handling, and evaluation.
AI agents are rapidly moving beyond chat-based assistants toward systems that can reason over data, interact with tools, and automate operational workflows. But applying these capabilities to network troubleshooting introduces unique challenges: network state is distributed across multiple systems, telemetry may be incomplete or stale, and an incorrect action can have significant operational impact.
This talk walks through the design and implementation of an AI agent for network troubleshooting, using a virtual router environment as a practical example. We will explore how to define the right problem for an agent, design its architecture, expose network capabilities through tools, and enable the agent to investigate routing and connectivity issues using real network state.
The talk will also focus on the engineering considerations that become critical when deploying agents in network operations, including tool selection, context management, deterministic workflows versus LLM-driven reasoning, guardrails, permissions, human approval, failure handling, and evaluation.










