I want to build an AI-powered enterprise investigation system and I want your advice on the best approach before I start building it.
Want to build an AI-powered enterprise investigation platform where users can ask natural-language questions across enterprise systems such as Jira, New Relic, Azure, AWS, etc. For now I'm starting with Jira + New Relic, with all integrations through MCP servers.
The goal is not just searching tickets; users should be able to ask arbitrary operational questions such as "Why did this incident happen?", "When did it happen?", "What caused it?", "Show related incidents", or "Give me correlated tickets including questions that require multiple systems and multiple steps of investigation.
I'm looking for advice from people who have built production-grade agentic systems. If you were starting this project from scratch, what approach would you choose for the agent/orchestration framework, MCP/tool selection, multi-step investigation and reasoning, evidence collection, and handling complex cross-system questions? What are the biggest challenges or failure modes I should expect, and what architectural decisions would you make differently to keep the system reliable, scalable, and maintainable as I add more enterprise systems? I'm deliberately not specifying my preferred framework or architecture because I want unbiased recommendations before I continue building.