Search in which an AI system plans and carries out multiple information-gathering steps using its available tools.
Agentic search uses an AI system to plan and carry out a sequence of information-gathering steps toward a goal. It may break a broad question into smaller searches, inspect results and adjust subsequent queries. The exact behavior depends on the application and the tools available to it.
Search and external action are different capabilities. A system that can compare options does not necessarily have access or authorization to book, buy, send messages or change records. Those steps require appropriate tools, permissions and any required approval. The term agentic does not remove those boundaries.
Multi-step retrieval can help address a complex question, but more steps do not guarantee better evidence. The system can pursue a mistaken interpretation, retrieve stale information or overlook an important constraint. A useful workflow keeps its sources and relevant actions inspectable.
Takeaway: Define the goal, access limits and approval points before using an agent for business work. Distinguish researched options from actions that have actually been authorized and completed.
Example
A business asks an assistant to compare venues against capacity and accessibility requirements. The assistant may gather public information and identify questions needing confirmation. Contacting venues or placing a deposit is a separate step, not an automatic consequence of the research request.
What to check
Check the final evidence against the original requirements and confirm important availability directly. Require appropriate authorization for consequential actions. Keep a clear record of completed actions rather than treating a plan or recommendation as an executed transaction.
For more detail, see Quickstart: Agentic retrieval.