How Deep Research Agents Work: Orchestrators, Parallel Sub-Agents, and Cited Synthesis
Inside deep research architectures across OpenAI, Gemini, and Claude: planning supervisors, parallel search sub-agents, and verifiable claim attribution.
On this page
In early 2025, frontier AI labs introduced a new class of long-horizon AI capabilities: Deep Research. Unlike standard chatbot prompts that return answers in two seconds, or basic Retrieval-Augmented Generation (RAG) that retrieves top-k chunks from a vector database, Deep Research operates as an autonomous, multi-agent research team.
When given an ambiguous, high-scope prompt (such as “Analyze the supply chain risks of sodium-ion battery manufacturing across Europe through 2030”), Deep Research runs for 10 to 30 minutes. It conducts hundreds of distinct web searches, parses PDF whitepapers, cross-references corporate balance sheets, resolves conflicting statistics, and compiles an exhaustive, cited report spanning dozens of pages.
Building a system that operates reliably over hundreds of autonomous tool calls requires moving beyond single-turn reasoning models to hierarchical multi-agent orchestration architectures.
+--------------------------------------------------------------------------+
| Hierarchical Deep Research Architecture |
+--------------------------------------------------------------------------+
| User Research Request |
| | |
| v |
| +----------------------------------+ |
| | Lead Orchestrator / Planner | |
| | (Decomposes scope & dependencies)| |
| +----------------------------------+ |
| | |
| +-------------------------+-------------------------+ |
| | | | |
| v v v |
| +---------------+ +---------------+ +---------------+ |
| | Search Agent | | Search Agent | | Analyst Agent | |
| | (Sub-Topic A) | | (Sub-Topic B) | | (Code & Data) | |
| +---------------+ +---------------+ +---------------+ |
| \ | / |
| +------------------------+------------------------+ |
| | |
| v |
| +----------------------------------+ |
| | Report Synthesizer Agent | |
| | (Outline, Themes, Narrative) | |
| +----------------------------------+ |
| | |
| v |
| +----------------------------------+ |
| | Citation & Fact-Checking Judge | |
| | (Maps every claim to verified URL| |
| +----------------------------------+ |
| | |
| v |
| Final Verified Cited Report |
+--------------------------------------------------------------------------+
Step 1: Upfront Scoping and Strategic Planning
The single largest predictor of research failure is ambiguous initial scope. If an agent begins executing searches against an underspecified prompt, it suffers from context drift: exploring irrelevant tangents and exhausting token limits.
Frontier providers solve planning through two primary paradigms:
+------------------------------------+-------------------------------------+
| Provider Implementation | Planning Strategy |
+------------------------------------+-------------------------------------+
| OpenAI Deep Research | Interactive Clarification Questions |
| Google Gemini | Structured Blueprint Proposal |
| Anthropic Claude | Lead Planner & Parallel Fan-Out |
+------------------------------------+-------------------------------------+
- Interactive Clarification (OpenAI): Before making a single web call, the system prompts the user with 2-4 clarifying multiple-choice or short-answer questions (such as desired geographic focus, target depth, or key competitors to include).
- Structured Blueprint Proposal (Gemini): The model decomposes the query into a multi-step research outline, identifying primary axes, data targets, and expected sources. The user can review, edit, or approve the blueprint before execution begins.
Step 2: The Orchestrator and Structured Delegation
Once the research blueprint is locked, the Lead Orchestrator Agent takes control. The orchestrator does not perform web searches directly. Instead, it functions as a research director:
{
"subagent_id": "search_worker_battery_patents",
"task": "Retrieve granted European patents for sodium-ion cathode binders from 2022 to 2025",
"constraints": {
"max_search_queries": 15,
"allowed_domains": ["epo.org", "uspto.gov", "sciencedirect.com"],
"time_budget_seconds": 180
},
"required_output_schema": {
"patent_number": "string",
"assignee": "string",
"filing_year": "number",
"raw_snippet": "string",
"source_url": "string"
}
}
By passing strictly typed JSON payloads with explicit search limits and domain constraints, the orchestrator prevents worker sub-agents from hallucinating unbounded search loops.
Step 3: Sandboxed Sub-Agent Tool Execution
A critical architectural principle across OpenAI, Gemini, and Claude is that agents never touch the internet directly.
Sub-agents emit structured tool calls (execute_search, fetch_page_markdown, run_python_analysis). The surrounding application runtime intercepts these calls, executes them within secure, sandboxed environments, and returns the parsed results to the sub-agent’s local context:
[Sub-Agent LLM] ===(Tool Call: search_web("sodium-ion CATL patent"))===> [Host Runtime]
|
v
[Sub-Agent LLM] <===(Filtered Markdown & Citations (3,000 tokens))=== [Search Engine]
- Markdown Scraping: Full web pages containing megabytes of HTML, advertising scripts, and navigation bars are stripped into clean, semantic Markdown.
- Dynamic Reformulation: If an initial search yields zero authoritative hits, the sub-agent analyzes why the query failed, reformulates search keywords, and executes a second targeted pass.
- Short-Term Context Isolation: Each sub-agent maintains its own isolated context window. Once its topical investigation concludes, it produces a condensed summary with source citations, discarding the hundreds of raw search pages. This prevents the primary orchestrator’s context from overflowing.
Step 4: Synthesizing the Narrative and Enforcing Citations
The final stage of deep research is assembly. Unifying 50 intermediate research notes into a cohesive 20-page document introduces severe hallucination risks: models often merge statistics from different years or attribute conclusions to the wrong company.
Frontier architectures decouple report drafting into two dedicated roles:
+---------------------------------------------------------------+
| 1. Report Synthesizer Agent: |
| - Aggregates thematic clusters across all sub-agent notes. |
| - Structures narrative headings, comparisons, and tables. |
| - Drafts long-form prose with placeholder source markers. |
+---------------------------------------------------------------+
|
v
+---------------------------------------------------------------+
| 2. Citation Verification Judge: |
| - Extracts every assertion of fact, number, or quote. |
| - Verifies claim against source URLs and raw text tokens. |
| - Replaces unverified assertions with explicit caveats. |
| - Injects verified Markdown hyperlinks [1](url). |
+---------------------------------------------------------------+
By separating synthesis from verification, Deep Research systems achieve the factual accuracy required for executive decision-making, patent analysis, and financial due diligence.
Frequently asked questions
What distinguishes a Deep Research agent from standard retrieval-augmented generation (RAG)?
Standard RAG executes a single vector lookup and generates an answer in seconds. Deep Research agents run autonomous multi-turn loops for 15 to 30 minutes, planning sub-queries, running hundreds of web searches, evaluating source credibility, and cross-synthesizing long-form reports.
How do orchestrator agents coordinate parallel research workers?
The orchestrator decomposes a broad research question into independent topical sub-tasks, assigns structured payloads with search permissions to parallel sub-agents, aggregates intermediate findings, and resolves dependency bottlenecks.
Why do Deep Research systems employ a separate citation agent during synthesis?
Generative models frequently hallucinate or misattribute quotes when summarizing long contexts. A dedicated citation verification agent matches every factual claim against the exact URL and extracted text snippet before report finalization.
What is the key difference between OpenAI and Gemini planning approaches?
OpenAI engages in multi-turn interactive clarification questions with the user before committing to search execution, whereas Gemini drafts a structured multi-step research blueprint upfront and requests explicit user sign-off.



/* Comments */
Comments are offline right now — we reconnect automatically, nothing is lost.