What the heck: Agentic AI???

I am the IT Manager. I am close with the development team as I support them quite a bit and still personally handle all CI/CD and Infra/Devops work. Data source is quite a general term. Agents grab context from numerous sources. Some of which are internally hosted. Examples being knowledge base (self hosted IT Glue), asset management, internal app, vpn. But a substantial amount of functionality is from integration with our IdP, CRM, random software different departments use, our HRIS. These "data-sources". Are obviously external. But these connections allow the agent to perform a lot of actions. I have also developed a lot of small algorithms/workflows to perform actions inside our internal apps, or through external apps with an open API. There can definitely be oddities in how the agent performs. LLMs are inherently non deterministic systems. So when performing deterministic actions, there are some important steps to take. Let's say you are developing a workflow for your agent to perform. This workflow takes an end user query and turns this into multiple system actions. It is important to spend substantial time asking the question you assume the user will, in every way possible. If you have an ITSM/Helpdesk tool where you can export prior tickets, this helps a lot. But doing this allows you to see and test how your agent would respond in the field. Let's go back to the probabilistic nature of LLMs. Your goal is to create funnels. Large funnels that capture a broad range of user queries on one end, determine intent and then what action should be performed, then perform the action. Reasoning models are quite good at this task. If you do this correctly and handle edge cases as a programmer would, agent oddities become close to a non issue. In terms of the oddities that can get past this implementation phase, our system has a record of every agentic action taken and connected to what user query that started it. If we see weirdness, it is easy to troubleshoot and temporarily disable that workflow. It is important to note that many actions the agent wants to take can require helpdesk approval, my approval, their manager's approval, and a mix of both. I can speak about this system at depth. I have barely scratched the surface and this has been my project for the better part of the last year. But If you're going to ask questions, try to make them specific.