LlamaIndex Agents

LlamaIndex started as a way to let a model answer questions about your own documents. LlamaIndex Agents is what happens when that retrieval machinery stops being the whole application and becomes one tool among several that an agent decides when to use.

Retrieval as a tool
The distinctive move is wrapping a query engine — an index over a document set, with its retriever and response synthesiser — as a callable tool with a description. The agent then reasons about which corpus to consult, or whether to consult one at all, instead of forcing every request through a single retrieval step. A support agent holds one tool per knowledge base, a finance agent one per annual report, and routes a question to the right one; a multi-part question is split into sub-questions answered against different indexes and then combined.

The loop and the workflow
Above the tools sit the usual agent forms — a function-calling agent for models with native tool use, a ReAct agent for those without — plus an event-driven workflow layer for multi-step and multi-agent runs, where steps emit and consume events and one agent hands control to another. Hundreds of ready-made tool integrations exist, so a corpus, an API and a database sit side by side in the same toolset.

Why it matters
Agents fail on knowledge they do not have, and retrieval is the standard fix. Treating the index as a tool rather than a pipeline is what lets an agent ground a long task in several document sets, deciding at each step which one the current question needs.