Beyond Intent Detection: Multi-Intent Reasoning And Execution In Agentic Observability Systems

Abstract

Multi-intent understanding is a fundamental requirement for AI assistants operating in complex, real-world domains such as observability and incident investigation, where user queries routinely encode multiple interdependent tasks within a single utterance. Existing approaches treat multi-intent detection as multi-label classification—identifying intent labels independently without modelling the dependencies, execution ordering constraints, or shared context that govern how compound queries must be resolved. This framing is insufficient for agentic systems that must translate user intent into coordinated tool-driven workflows. This paper introduces an execution-aware multi-intent framework that reframes multi-intent understanding as a planning and orchestration problem. User queries are parsed into structured intent sets with shared entity context, transformed into an intent graph that encodes dependency and ordering constraints as typed directed edges, compiled into an execution plan through topological analysis, and executed through a deterministic orchestration layer that dispatches tool calls, routes results across intents, and recovers from partial failures. Evaluation on a benchmark of 1,240 hybrid synthetic and realistic multi-intent observability workflows demonstrates significant improvements over single-intent, independent multi-intent, and LLM-only implicit orchestration baselines across task completion rate, dependency handling accuracy, and robustness under partial failures. The results establish that modelling execution semantics—not merely intent labels—is the critical architectural requirement for production-grade multi-intent agentic AI systems. The framework and its implications extend beyond observability to any domain where natural language user queries must drive coordinated multi-step execution.