1 of 5

DEEP RESEARCH AGENT�EXAMPLES

GEORGI NAYDENOV

2 of 5

SINGLE AGENT WORKFLOW

User

Simple Search Agent

The agent reviews whether it can answer directly or call Perplexity

The agent stores the conversation in its short-term memory.

The model generates a response.

User sends a search request

Send message

3 of 5

SEARCH ORCHESTRATION

Search Planner

The agent processes its system prompt.

Breaks down the user request into specific subtasks

Outputs subtasks in JSON format

User�(Replit Web UI)

Webhook

Receive POST Payload to n8n

/POST

Search Orchestrator

Ingests JSON subtasks.

Task execution is parallelized.

Call Search Worker

Aggregate the JSON Results

LLM�Output Parser�

LLM�Search Worker�Query Results Store

AI Summarizer

Read results JSON

Synthesize insights

Generate Markdown report

LLM

Transform MD to HTML

POST response payload

Render in Replit UI

Send Report

Ship to UI

Search Worker (executed by Orchestrator)

​

  • Build query from a subtask
  • Call https://api.exa.ai/search
  • Receive raw results

EXA API REQUEST

HTTP POST

​

  • Extract top items
  • Flatten fields
  • Filter noise

SPLIT OUT

Normalize

​

  • Title, url, snippet
  • Source, confidence
  • Emit results[]

EDIT FIELDS

Shape JSON

  • Send structured results
  • Attach subtask id

RETURN TO ORCHESTRATOR

Results JSON

Aggregate results

4 of 5

MULTI-AGENT RESEARCH SYSTEM

Form Submission�(n8n form)

Customer Support Agent

Clarify scope

  1. Parse initial request
  2. Ask clarifying questions
  3. Confirm report generation started
  4. Emit normalized research scope

Open AI Model

Simple Memory

Structured Output Parser

  1. User enters task
  2. Upload optional context

Submit Research

  1. Tweak the initial JSON
  2. Override defaults

Edit Fields

Lead Agent (Orchestrator)

Plan, coordinate 3-7 Search Subagents

  1. Analyze scope & produce subtasks
  2. Dispatch parallel subagent processes
  3. Monitor progress & reflect
  4. Aggregate findings for the copywriter agent

Open AI Model

Simple Memory

Think/Parser

Search Subagent

Copywriter Agent

Synthesize, format & deliver

  1. Analyze scope & produce subtasks
  2. Dispatch parallel subagent processes
  3. Monitor progress & reflect
  4. Aggregate findings for copywriter
  5. Create final report

Open AI Model

Think

Think/Parser

Structured Output Praser

  1. POST Markdown
  2. Receive final PDF

HTTP REQUEST (MarkdownMaster)

  1. Attach the PDF to response
  2. Expose download link to Google Drive API

UPLOAD FILE

Search Subagent

Search the Internet

  1. Query Brave for 8-12 results
  2. Visit 3-9 pages for a deep read
  3. Call the Fetch Website tool to extract
  4. Summarize & tag key findings
  5. Return JSON back to Lead Agent

LLM

Brave Search Agent

  1. Triggered once the Lead
  2. Agent makes a call to the Search Subagent

Workflow Trigger

  1. Search Agent Trigger
  2. Website Fetch trigger

Switch Function

Website Fetch

Think

  1. Call Scraping Ant API
  2. Send Findings to LLM

HTTP GET REQUEST

  1. Review findings and create learning notes
  2. Format Findings

LLM Chain

  1. Review LLM Chain Content
  2. Condense Findings and send to Lead Agent

Gemini 2.5 Lite

Fetch a Website

Output findings in JSON

5 of 5

AGENT SWARM — HIGH-LEVEL ORCHESTRATION

Customer Support Agent

Clarify scope

  1. Parse initial request
  2. Ask for missing details
  3. Capture goals and constraints
  4. Send scope JSON to Planning Agent

Open AI Model

Simple Memory

Structured Output Parser

  1. User enters task
  2. Upload optional context

Submit Research

  1. Tweak the initial JSON
  2. Override defaults

Edit Fields

Planning Orchestrator

Create Research Plan

  1. Review scope JSON
  2. Decompose into topics/tasks
  3. Create the initial research plan

Deep Seek R1

  1. User enters task
  2. Upload optional context

Split and Merge

  1. Research plan to G.Sheets
  2. Send out request to Slack

Google Sheets Connector

Structured Output Parser

Slack Approval

Human-in-the-loop

Approve

Deny

Request changes

  1. Send plan to requester DM
  2. Capture feedback
  3. Persist decision

Approve

Request Changes

Reviewer Loop�If changes are requested

  1. Capture feedback in Slack (n8n form)
  2. Planner updates plan JSON
  3. Re-submit for approval
  4. Audit trail preserved (n8n)

Push to Agent Swarm

Agent Swarm (Plan Approved)

​

  1. Search API (Tavily)
  2. Split Out Results
  3. Separate items for each URL
  4. Aggregate Results
  5. Merge Results
  6. Extract HTML Content
  7. Combine all Inputs
  8. Send to Google Sheet

​

Sources Agent

Compiles the final list of agents

  1. Read the Google Sheets content from Agent Swarm
  2. Compile sources for each section’s content
  3. Input sources to the Google Sheets document

LLM

Google Sheets

Merge Function

ToC Agent

Compile the Table of Contents

  1. N8N Code Node to combine multiple fields into a single field with new lines
  2. Title, Intro, ToC, Chapter Formatting
  3. Generate PDF Trigger

LLM

Javascript Function

Google Sheets

Sheets-to-PDF

​

  • Sheets to PDF
  • APITemplate.io API call
  • Download PDF via GET request
  • Send Report to Email

​

Form Submission�(n8n form)