What Is Agentic Search?

Agentic search is search where an AI agent interprets intent, plans a retrieval path, uses the right sources or tools, and synthesizes an answer instead of only returning ranked results. Whether it is reliable comes down to context engineering: the deliberate design of what data, tools, and memory the agent is allowed to reach, and how precisely that context is curated before the model reasons over it. In analytics, the useful version does not search raw tables and hope the model infers business meaning. It searches trusted context - governed metrics, dimensions, joins, and access rules - so the answer is grounded in the same definitions people use in BI, embedded analytics, and operational workflows.

A working definition of agentic search

Agentic search is a retrieval workflow controlled by an AI agent. The agent receives a goal, works out what information is needed, chooses where to search, retrieves evidence, checks whether the evidence is sufficient, and then returns an answer, artifact, or next action. A simple query like "show the latest invoice policy" might take one retrieval step. A broader question like "why did enterprise expansion slow last quarter?" needs planning, multiple searches, governed metrics, and a follow-up path.

That planning loop is the difference. Traditional search ranks documents. Semantic search improves ranking by matching meaning, not just exact words. Agentic search can use both, but it wraps them in a workflow: understand the request, decompose it, retrieve what matters, reason over results, and retrieve again when the evidence is incomplete. Search becomes less like a list of candidates and more like a task path an agent can execute.

Agentic search vs. semantic search vs. RAG

These terms overlap, but they are not interchangeable.

Keyword search matches terms. It is fast and explainable, but it struggles when the user's language does not match the indexed content. Semantic search uses embeddings or other semantic signals to retrieve results that are close in meaning, so "customers at risk" can find churn-related content even when those exact words are missing.

RAG, or retrieval-augmented generation, retrieves context and gives it to a model before it generates an answer. That is a useful pattern, but many RAG systems are still one-shot: retrieve top chunks, answer from those chunks, stop. If the question needs two hops, a different source, or a tool call, the system often misses the gap.

Agentic search turns retrieval into a loop. The agent can ask, "What do I need to know first?", run one search, inspect the result, use it to form a better second query, call an API, and only then answer. In a data setting, that is close to the work described in AI agents for data analysis: the agent is not just finding text, but choosing a path through governed information.

Agentic search is a context engineering problem

The interesting engineering in agentic search is not the loop itself — it is context engineering: the strategic design of what context gets fed to the LLM. Data, tools, memory, and prior results all compete for a limited, high-stakes context window, and the quality of an agent's answer is bounded by what it is allowed to retrieve and how precisely that material is curated.

This reframes the goal. The aim is not to give the agent more context; it is to give it the right context — only the most relevant, high-precision information, with the ambiguity already removed. Flooding the model with raw documents, every table in a warehouse, and half-matching search hits makes answers less reliable, not more, because the agent has to guess which fragment is authoritative. Good context engineering does the curation up front: it decides which sources are trusted, resolves what a term means before the agent sees it, and enforces who is allowed to see what.

For analytics that discipline is not optional. The context an agent searches has to carry business meaning and governance with it, or the search path quietly drops both. The rest of this article is about building that trusted, high-precision context space.

Why agentic search is harder for analytics

Enterprise document search can tolerate some ambiguity. A policy answer may cite the wrong paragraph and still be reviewable by a human. Analytics has a narrower error budget because the answer is a number, a segment, a trend, or a decision someone will act on.

Raw data systems are not written in business language. A warehouse table might contain orders, subscriptions, refunds, trials, seats, usage events, and slowly changing account attributes. None of those names tells an agent what "ARR" means, whether refunds are included in revenue, which join path preserves grain, or which tenant rows the current user may see.

Point an agentic search workflow at that raw schema and it can retrieve relevant-looking pieces while still building the wrong answer. It may find a dashboard title, a SQL snippet, and a table column that all mention revenue, then combine them without knowing which one is certified. That is the same failure mode behind raw text-to-SQL: the model re-derives joins and metric logic every time, so the same question can return different numbers.

The semantic layer is the trusted search space

For analytics, a semantic layer is the trusted search space an agent needs — and the place context engineering actually happens. It defines metrics, dimensions, join paths, and access rules once, then serves that curated model to downstream tools. Instead of searching raw tables, the agent discovers certified business objects and asks for them by name, so the context it retrieves is already high-precision and unambiguous.

That changes what retrieval means. The agent is not asking, "Which table looks like it has revenue?" It is asking, "Which governed revenue metric matches this question, and what dimensions and filters are allowed for this user?" The semantic layer compiles that request into SQL against the warehouse, applies row- and role-level rules before the query runs, and can use caching or pre-aggregations for common query shapes.

This is why semantic layers for AI agents and agentic search belong together. The search layer helps the agent find the right context. The semantic layer makes that context precise and safe to use. One without the other is incomplete: retrieval without governed definitions is brittle, and governed definitions without an agent-friendly discovery path are hard for users to reach in natural language.

What good agentic search needs

A production agentic search system needs more than a vector index and a chat box. The important capabilities are structural, and most of them are context engineering in practice:

• Intent planning. The agent should identify whether the request is a lookup, a multi-step question, or an action-oriented workflow.
• Source routing. It should know when to search documents, when to query metrics, when to call a product API, and when to ask for clarification.
• Iterative retrieval. It should inspect intermediate results and search again when evidence is missing or contradictory.
• Curated, governed context. It should retrieve only the most relevant, high-precision context from approved sources, with permissions enforced before results reach the model.
• Traceability. It should show which metrics, dimensions, filters, documents, or tool calls produced the answer.
• Consistent interfaces. Agents and applications need stable ways to reach the same governed model, including SQL, REST, GraphQL, and MCP.

The last point matters because agentic search is not only an internal productivity feature. Product teams also want AI analytics inside their own applications, where customers ask questions about their own data. Embedded analytics raises the bar: every search and query path must be tenant-aware, fast, and auditable.

Where Cube fits

Cube is the agentic analytics platform built on a semantic layer. Its open-source foundation, Cube Core (Apache 2.0), defines metrics, dimensions, joins, and access rules, then serves them through SQL, REST, GraphQL, and MCP. The Cube platform adds AI agent interfaces, workbooks, dashboards, embedded surfaces, multi-tenancy, and managed performance on top of that foundation.

That makes Cube the context engineering layer for agentic search over data. The agent needs a place to search for trusted analytical context: available metrics, allowed dimensions, joins, filters, and user-specific permissions. Cube's AI context layer curates exactly that — the governed model, resolved to certified definitions and scoped to what each user may see — so the agent retrieves only the most relevant, high-precision context instead of raw schema. The same definitions can then power internal BI and embedded customer-facing analytics.

This does not replace the warehouse or dbt. The warehouse still stores and computes over the data in Snowflake, BigQuery, Redshift, or Databricks. dbt still models and transforms data upstream, and Cube can read dbt models. The semantic layer sits above those systems and engineers the business context an agent needs to search, query, and answer consistently.

How to evaluate agentic search

When you evaluate agentic search for analytics, do not stop at whether the demo answers a natural language question. Ask five questions:

1. Does the agent retrieve governed business definitions, or does it infer meaning from raw schema and dashboard text?
2. Are row-level, role-based, and tenant permissions enforced before the model sees results?
3. Can the agent show which metric, dimension, filter, and time range produced the answer?
4. Does the same model serve internal BI and embedded analytics, or are there separate definitions?
5. Can the retrieval path use standard interfaces such as SQL, REST, GraphQL, and MCP?

If the answers are weak, the system may still be useful for document discovery, but it is not ready to answer analytical questions in production. The bar for analytics is higher: search has to find the right context, and the context has to carry business meaning and governance with it. The agent can plan and retrieve; disciplined context engineering on a semantic layer keeps the answer precise, governed, and explainable.