What is agentic RAG?
Last reviewed 7 October 2026
Agentic RAG is retrieval-augmented generation in which an AI agent controls the retrieval step, deciding what to search for, which sources to use and whether to search again before it answers. The survey by Singh and colleagues describes it as embedding autonomous agents into the RAG pipeline. Standard RAG, by contrast, retrieves once and answers from whatever came back.
Also called: agentic retrieval-augmented generation, agentic retrieval, agent RAG.
Why Agentic RAG matters
Standard RAG fails quietly on certain questions. If the first search returns the wrong passages, the model answers confidently from them. Questions that need two lookups, where the second depends on the first, cannot be answered by one similarity search at all. Agentic RAG addresses both by letting the model inspect what it retrieved and try again.
It is not free. Each extra reasoning step and search is another model call, so answers take longer and cost more. Most question-and-answer systems over a single, well-organised corpus do not need it. The full guide linked below works through when it pays for itself.
How it works
The survey names four agentic patterns found in these systems: reflection, planning, tool use and multi-agent collaboration. In practice that usually means a loop. The agent reads the question, plans one or more searches, calls retrieval tools, judges whether the results are enough, and either rewrites the query and searches again or writes the answer.
Retrieval tools can be anything the agent may call: a vector index, keyword search, a SQL query or a live API. Microsoft’s agentic retrieval in Azure AI Search is one vendor’s version of the planning part: an LLM can break a complex query into focused subqueries, run them in parallel, rerank and merge the results. On Microsoft’s page read on 7 October 2026, LLM-based query planning was marked as preview.
Example: a two-part policy question
An HR adviser asks: “Can an employee on a fixed-term contract in our Dublin office carry over unused leave?” A single search for that sentence returns general leave policy pages. The agent notices that none mention fixed-term contracts, searches the contract templates for the term, finds the relevant clause, then searches the Ireland policy supplement for carry-over rules.
It answers with both sources cited. If the two documents conflict, it says so and names them rather than picking one. The agent never had more access than its search tools gave it.
How Swfte relates to it
Designed for
Swfte does not sell agentic RAG as a packaged feature. Swfte Studio is designed for assembling the loop yourself: an agent with retrieval tools, a grading step and an answer step, where each step can use a different model through Swfte Connect. Connect provides the routing, fallback between providers and per-call cost record that a retrieval loop needs.
Swfte Cortex includes on-device knowledge bases with retrieval over your own files. For the architecture, the costs and a build order, read the agentic RAG guide rather than this short definition.
- Swfte Studio
- Swfte Connect: Per-step model routing
- Swfte Cortex: On-device knowledge bases
Related terms
- Retrieval-augmented generation (RAG)
Retrieval-augmented generation (RAG) is a method in which a system first retrieves relevant passages from a collection of documents and then gives them to a language model as context for its answer.
- Agentic search
Agentic search is a search process in which an AI model decides when to search, what to search for and whether to search again, instead of running one fixed query.
- AI agent
An AI agent is a software program that uses an AI model to decide what to do next and then acts through tools to reach a goal it was given.
- Agentic AI
Agentic AI is a style of AI system that works toward a goal with limited supervision, choosing its own next steps and calling tools to act on other systems.
Common questions
- How is agentic RAG different from RAG?
- Standard RAG runs a fixed pipeline: one search, then an answer. Agentic RAG lets an agent decide whether the search worked, rewrite the query, choose another source and search again. The retrieval machinery underneath is often the same; the difference is the decision loop on top of it.
- When is agentic RAG worth the extra cost?
- When questions need more than one lookup, span several sources with different shapes, use different vocabulary from your documents, or need live data from an API. For single-topic lookups in one well-indexed corpus, standard RAG is usually faster, cheaper and easier to test.
- Does agentic RAG stop hallucinations?
- No. It reduces one cause, answering from the wrong passages, by letting the agent check and search again. The model can still misread a correct passage or fill a gap. Require citations, tell the agent to say when it found nothing, and test against questions with known answers.
- Is agentic RAG the same as MCP?
- No. Agentic RAG is a retrieval strategy. The Model Context Protocol is a standard way for tools and data sources to present themselves to AI applications. An MCP server that exposes a search tool can be one of the retrieval tools inside an agentic RAG loop.
Sources
Definitions on this page were read on the sources below on 7 October 2026. Where sources define the term differently, the page says so. The full glossary lists more terms.