Agentic Search

Definition

Agentic search refers to search systems that go beyond simple retrieval to plan, decompose, execute, verify, and refine — functioning more like engineering teams pursuing goals under uncertainty than simple lookup functions.

The key distinction from traditional search: instead of a single-pass query → results interaction, agentic search operates through iterative, multi-step workflows managed by AI agents.

Core Properties

  • Iterative uncertainty reduction — each step should be evaluated by how much uncertainty it removes per unit of cost
  • Decomposition — complex queries broken into subtasks, delegated to specialized tools
  • Verification — intermediate results checked and refined
  • Evaluation-driven completion — system stops when marginal gains no longer justify cost

The Scope-Cost-Quality Triangle

Adapted from classic project management:

DimensionMeaning
ScopeBreadth of problem space explored
CostToken/computation spending
QualityCorrectness, completeness, confidence

Three operational strategies:

  1. Reduce scope (fixed cost & quality) — narrow focus, early stopping; default behavior
  2. Increase cost (fixed scope & quality) — deep research mode; broader exploration
  3. Sacrifice quality (fixed scope & cost) — quick overview mode; skimming + summarization

Relationship to Agile Development

AgileAgentic Search
Product ownersSearchers
EngineersAgents
Sprints (days)Iterations (seconds/minutes)
Tests define doneEvaluation defines done

Normalization Note

“Agentic retrieval”, “agentic search”, “agent-based retrieval”, and “AI agent search” all refer to the same paradigm. This note uses Agentic Search as the canonical term.

  • Retrieval Pipeline — the multi-stage pipeline that agentic systems orchestrate

  • RAG — foundational architecture that agentic search extends

  • Query Understanding — understanding intent is especially critical in agentic workflows

  • Context Engineering — agentic search is ~80% of context engineering; the curation arrow maps to search tools

  • Search-R1 — RL-trained multi-turn search-and-reasoning agent; concrete implementation of agentic search

  • Reinforcement Learning for Search — training paradigm that enables autonomous search strategy learning

Articles

People