C# AI agent uses Tavily to research .NET errors

A .NET Error Research Agent forces AI to consult external sources like GitHub, StackOverflow, and Microsoft documentation before responding to exceptions, cutting off hallucinated fixes [Dev.to]. The agent, built in C# using Microsoft's Semantic Kernel and Azure OpenAI, integrates the Tavily Search API to mandate real-time research for every error query. It runs on a strict workflow: receive exception, trigger search, analyze results, then respond — with no path for bypassing external validation [Dev.to].

The core mechanism is a custom WebSearchPlugin that packages the full exception and stack trace into a Tavily query targeting GitHub issues, StackOverflow threads, and official docs [Dev.to]. Search depth is set to "advanced" with up to five results returned, including extracted answers. Only after this step does the AI generate analysis or code.

Responses follow a rigid structure: exception summary, root cause analysis, ranked possible causes (high/medium/low probability), explanation, recommended fix, and corrected code [Dev.to]. The agent targets common .NET pain points: dependency injection failures, NuGet package mismatches, Entity Framework misconfigurations, and ASP.NET Core runtime errors.

This approach disables the default AI behavior of answering from memory, instead treating the model as an analyst that must research and weigh evidence [Dev.to]. By constraining the AI to research-first behavior, the agent treats knowledge as dynamic, not frozen in weights. This rejection of the "smart model" paradigm is a template others should copy, particularly in domains where lived experience trumps memorization.