Agentic Search
How it works
1) User query intake: user asks a complex question ("Deep dive into chip market share 2024, TSMC vs Samsung vs Intel"). 2) Planning: LLM (with reasoning mode/thinking budget) generates an action plan: outline, sub-queries, source types, ordering. 3) Query decomposition: breakdown into 5-40 sub-queries. 4) Parallel/sequential search execution: agent runs queries (Google, Bing, Brave, ArXiv, Scholar) via tool APIs. 5) Result fetching: web scraping (rendered HTML or markdown), PDF extraction (arXiv papers), rate-limited API calls. 6) Reading and note-taking: LLM reads each result, extracts key quotes and facts. 7) Reflection loop: LLM assesses coverage, identifies gaps, generates follow-up queries. 8) Iteration: steps 3-7 repeated 3-10 times typically, up to 30+ for deep research. 9) Structured synthesis: hierarchical report planning with sections, integrating citations. 10) Citation formatting: every statement with a source (URL + quote + date + author), inline citations. 11) Output: comprehensive report (5-20 pages), interactive follow-up capability, ability to expand sections. Test-time compute: 1-30 minutes, costs $0.10-$5.00 per query.
Problem solved
Classical RAG (single-shot retrieve-then-generate) has three fundamental limitations: (1) dependence on single semantic query quality — if the user asks poorly, the retriever returns poor documents; (2) no planning — cannot decompose a complex question into components; (3) no reflection — the model cannot verify whether top-K contains enough info. Agentic search solves these problems via a reasoning + tool-use loop. For complex questions (research, comparison, analysis), it delivers 10-30x better accuracy than RAG. Enables use cases previously impossible: 10-30 minute deep research sessions with 40+ sources, comprehensive reports with sections, multi-perspective analysis, hard fact-checking with citations.
Components
Implementation
Evolution
Hyperparameters (configurable axes)
Quick answer (1-3 queries, <30s) vs standard (5-15 queries, 1-3 min) vs deep research (30-100+ queries, 10-30 min). OpenAI Deep Research: 10-30 min. Google Gemini Deep Research: 5-15 min. Kimi Researcher: 5-30 min. Defines UX + cost.
Web only (Google/Bing) vs academic (ArXiv, Scholar, PubMed) vs specialized (Bloomberg for finance, LexisNexis for law, Kaggle for datasets). Critical for domain expert use cases. Kimi Researcher supports 8+ tool types.
Strict (every statement must have a source) vs lax (major ones cited, minor without). Strict increases trust but reduces fluency. Perplexity: strict inline. Kimi: strict + widget breakdown.
Long research sessions consume 100k-1M tokens. Options: full context (long-context model like Gemini 3.1 Pro 1M), sliding window summarization, hierarchical compaction (Kimi K3 @300K trigger), episodic memory. Compaction quality drives long-run quality.
Parallel: faster but cannot adapt queries to results. Sequential: adaptive but slower. Hybrid: initial burst of 5-10 queries in parallel, then adaptive. Standard in production.
Perplexity: chat-style with citations. OpenAI/Gemini Deep Research: structured multi-page reports. Kimi Work: rich reports with widgets. Format affects perceived quality and use case fit.
Execution paradigm
Execution steps conditional on discoveries — when info is insufficient, additional queries; when it's enough, move to synthesis. Adaptive stopping criterion. Reasoning effort scaled by query complexity.
Parallelism
Parallel execution of sub-queries (dozens of fetches in a burst). Sequential phases: planning, reflection, synthesis. LLM decoding sequential (autoregressive). Parallelism at the user level (each session is independent).