Selector Al Agents Architecture

Selector AI Agents Architecture is built to interact with enriched data and gain insights through a natural language interface, supporting web portals and integrations like Slack and Teams. It transforms how users chat with their network to uncover critical information. The initial approach, conceived in 2023 at the advent of public LLMs, involved translating natural language questions into STQL (Selector Query Language), running a single query, and sending the results to an LLM for reasoning. This single-shot method proved insufficient for complex networking issues, which often require multiple data sources and iterative steps. Limitations included small context windows, the need for pre-seeded translation phrases, and challenges with accurate natural language understanding, such as misinterpreting named entity recognition (NER) for terms. Selector evolved to a multi-turn reasoning AI agent architecture, solving these early shortcomings. This system employs a React pattern, allowing agents to iteratively plan, execute tools, and observe results to reach a final answer. The architecture is three-tiered: a central orchestrator agent acts as the "general contractor," planning and coordinating tasks while maintaining conversational context. It dispatches requests to domain-specific worker agents—like firewall, load balancer, or cloud observability agents—which act as "specialized plumbers or electricians." These worker agents, which are also React agents, have a limited set of tools (the "tool belt") available via the Microservice Control Plane (MCP) and focus solely on answering their specific query before reporting back to the orchestrator. Selector emphasizes LLM agnosticism, offering connections to private, enterprise, or secure public LLMs (like Gemini with data non-training guarantees), and provides extensive auditability and traceability, logging every agent action and decision through OpenTelemetry and MongoDB. The platform supports two main integration patterns: using Selector’s orchestrator with customer-provided tools, or customers integrating their own agentic ecosystems with Selector's MCP toolset. This enables guided remediation, from creating maintenance windows and alert rules to executing Ansible playbooks and integrating with third-party workflow engines like Itential for configuration changes. Joby Rudolph demonstrated these capabilities by showcasing how an agent can diagnose unreachable applications by synthesizing data from synthetics, routing, anomalies, and cloud agents, providing both root cause analysis and recommended actions. He also showed a closed-loop remediation where an agent used an external Itential workflow to correct a typo in a device configuration, highlighting the system's extensibility and ability to foster trust by allowing human-in-the-loop validation for critical actions. The architecture's flexibility allows customers to define new agents and tools, leveraging Selector's underlying enriched data for comprehensive network observability and management. Presented by Joby Rudolph, Senior Distinguished Engineer. Recorded live at AI Field Day 8 in San Jose, California on May 13, 2026. Watch the entire presentation at https://techfieldday.com/appearance/selector-ai-presents-at-ai-field-day-8/ or visit https://TechFieldDay.com/event/aifd8/ or https://Selector.AI for more information.