Why do RAG chatbots stop at complex requests?
Build an enterprise RAG chatbot and the limits show up fast. It receives a question, searches documents, and generates an answer — this straight-line structure is a single-track pipeline. Faced with a request like “find last year's Q4 results and draft this year's strategy proposal,” a pipeline that only searches and answers has no path to carry out the work at all.
Why are loops and branches the core of intelligence?
When a convenience store is out of water, a person opens a map and finds the next store. A single-track pipeline says “not available” and stops. If a search turns up nothing, change the query and try again; if that still fails, switch to another tool — loops that backtrack when stuck and branches that change course with the situation are the core of intelligence, and designing them is orchestration.
The roles split three ways. LangChain provides the standard parts for talking to an LLM, LangGraph the control flow that draws loops and branches, and LangSmith the observability that lets you see inside the system. Only together do the three become one orchestration system.
What do you lose without observability? — Automating tax consulting
Observability is often neglected, but without it you cannot trace back the grounds of an AI's answer. Take tax consulting: a staff member spends an hour or two a day digging through tax-law provisions and a client's gifting history. Automating that work requires branching that moves between statute search and history lookup.
And because a wrong answer becomes a matter of legal liability, an audit log must remain of which provision and which history the answer was based on. Leaving that trail of grounds is exactly what observability is — and it is LangSmith's job.
What do non-developers own in this design?
The non-developer's share of this design is larger than expected. Data ownership — deciding what data goes in; drawing the business logic as a flowchart; building an evaluation set from 100 questions with model answers — all of it is domain knowledge, not coding. Once you can draw the flow, vibe coding will translate it into code.
That said, attaching orchestration to a job a single API call can finish is a waste. It is a tool to reach for when the work is complex enough to need loops and branches. The starting point of AX is not the tools, but the domain knowledge to draw the workflow yourself.