5K

Agent-Driven Research

Agent plans its own search queries, runs them across sources, reflects on gaps, and iterates until it can write a sourced report

established Updated 2026-10-13 By Nikola Balic (@nibzard) Based on Danny Tarlow, Connie Fan
View .md Edit on GitHub
Add to Pack
or

Saved locally in this browser for now.

Cite This Pattern
APA 
Nikola Balic (@nibzard) (2026). Agent-Driven Research. In *Awesome Agentic Patterns*. Retrieved October 14, 2026, from https://agentic-patterns.com/patterns/agent-driven-research
BibTeX 
@misc{agentic_patterns_agent-driven-research,
  title = {Agent-Driven Research},
  author = {Nikola Balic (@nibzard)},
  year = {2026},
  howpublished = {\url{https://agentic-patterns.com/patterns/agent-driven-research}},
  note = {Awesome Agentic Patterns}
}

Use when

  • Open-ended research question needs multiple search rounds
  • Answer needs synthesis across many sources

Avoid when

  • A single retrieval round answers the question
  • Token cost or latency budget is tight

Problem

Traditional research methods often lack the ability to adapt search strategies based on emerging results, limiting efficiency and potential discoveries. Complex research tasks require multi-round investigation, cross-source synthesis, and dynamic strategy adjustment that static retrieval systems cannot provide.

Solution

Allow AI agents to independently conduct the entire research process. Given a research question, the agent:

  • Creates its own search queries through dynamic planning.
  • Executes the searches across multiple sources.
  • Examines the data and evaluates quality/relevance.
  • Reflects on whether sufficient information has been gathered.
  • Adjusts its search strategy based on findings and gaps.
  • Repeats until satisfaction criteria are met.
  • Synthesizes findings into a comprehensive, well-sourced report.

The key mechanism is a self-reflective iteration loop: after each search cycle, the agent evaluates results against its research goals and autonomously determines the next exploration direction. Production implementations typically combine four subsystems: planning (task decomposition), memory (context + vector store), action (tool orchestration), and reflection (quality evaluation).

A common variant uses parallel multi-agent teams, where different agents simultaneously pursue different research angles and later synthesize findings.

Relationship to AI Web Search Agent Loop: AI Web Search Agent Loop applies the same iterate-until-satisfied loop to real-time web question answering, with an emphasis on query translation, search operators, and inline citations. This pattern is the general form: an end-to-end research workflow over arbitrary sources (web, databases, documents) with explicit planning, memory, and reflection subsystems that ends in a written, sourced report.

Example (flow)

flowchart TD A[Research Question] --> B[Formulate Search Query] B --> C[Execute Search] C --> D[Analyze Retrieved Information] D --> E{Sufficient Information?} E -->|No| F[Refine Search Strategy] F --> B E -->|Yes| G[Synthesize & Summarize Findings] G --> H[Present Results to User]

How to use it

  • Use for open-ended research requiring strategy adaptation and multi-source synthesis.
  • Use when tasks need explicit control flow between planning, execution, and fallback.
  • Define clear termination conditions (satisfaction-based or resource-limited).
  • Design for multi-source integration (web, databases, documents).
  • Combine with reflection loop for self-correction; with agentic RAG for document retrieval.

Trade-offs

  • Pros: Enables autonomous multi-round investigation; adapts strategy based on findings; produces comprehensive, well-sourced outputs; superior for complex analytical tasks.
  • Cons: Higher token cost — operators commonly estimate 5-10x versus single-round retrieval, though this is an estimate rather than a benchmarked figure; increased latency from multiple LLM calls; planning stability challenges; orchestration complexity adds states to debug.

References

  • Primary source: 36kr interview with Connie Fan and Danny Tarlow (Gemini team) on agentic coding and search — Tarlow on how developers work in a large codebase: "You would use code search, look at the file structure, click around, read some code" and "The agentic model mimics this approach." The interview describes a dual-track approach: longer context plus an agentic programming model that searches and reasons autonomously.

  • ReAct (Reasoning + Acting): Foundational pattern establishing the Thought → Action → Observation loop; Yao et al., Princeton University & Google Research, ICLR 2023

  • "The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery": Automated research lifecycle from idea generation to manuscript writing; Sakana AI + Oxford + UBC, arXiv:2408.06292, 2024

  • "From AI for Science to Agentic Science: A Survey on Autonomous Scientific Discovery": Survey of autonomous research agents and scientific discovery systems; Shanghai AI Lab, arXiv:2508.14111, 2025

  • Tongyi DeepResearch: Open-source agent-driven research system with 60% inference cost reduction; Alibaba Tongyi Lab, arXiv:2510.24701, 2025; code at Alibaba-NLP/DeepResearch

  • Building Effective Agents (Anthropic) — orchestrator and agent-loop patterns behind production research agents