websearch-agent — Deep Web Researcher

Overview

The websearch-agent is a specialized subagent for finding high-quality, non-academic information from the web. It conducts multi-query searches, extracts and condenses webpage content, and discovers relevant images — all while maintaining strict output size limits to prevent context overflow in downstream agents.

Dictionary-Based SubAgent Definition:

websearch_subagent = {
    "name": "websearch-agent",
    "description": "Conducts deep web research and extracts webpage content, returning condensed summaries (not raw dumps) with relevant images.",
    "system_prompt": "...",  # Full prompt below
    "tools": [internet_search, extract_webpage],
    "model": subagent_model  # Default: gemini_3_flash
}

When the Main Agent Should Invoke This Subagent

• •Tier 3 (Deep Research) — Discovery phase, called in parallel with academic-paper-agent
• •Any time the main agent needs comprehensive web research beyond a simple internet_search call
• •When the research requires multiple searches, webpage extraction, and synthesis

Invocation Pattern:

task(name="websearch-agent", task="Research recent developments in quantum computing, focusing on error correction and practical applications.")

Tools

Research Process

Step 1: Query Decomposition

Break down the user's research goal into 3–5 specific search queries.

Step 2: Web Search

Use internet_search to find relevant URLs and images for each query.

Step 3: Content Extraction

Use extract_webpage to retrieve full content from top 3–5 URLs ONLY.

Step 4: Immediate Summarization (CRITICAL)

After each extraction, immediately condense the content into 3–5 bullet points:

[10,000 word raw content] →
"From [Source]:
 - Key fact 1
 - Key statistic 2
 - Important quote 3"

Step 5: Synthesis

Combine all summarized findings into the final structured output.

Context Management Rules (CRITICAL — Prevent System Crash)

⚠️ These rules are non-negotiable:

Image Handling

The internet_search tool returns images in its response. The agent MUST:

1. •Review returned images from each search
2. •Select top 2–3 most relevant, high-quality images that:
   • •Directly relate to the research topic
   • •Come from reputable sources (avoid ads, low-res, or irrelevant images)
   • •Add visual value (diagrams, charts, infographics, product images)
3. •Include selected images using Markdown syntax:
   markdown

   ![Descriptive Caption](image_url)
   

Example:

![Architecture diagram of multi-agent systems](https://example.com/diagram.png)

Output Format (STRICT — MAX 800 WORDS TOTAL)

## Summary of Findings
A comprehensive 2–3 paragraph summary (200–300 words).

## Key Evidence
- Fact 1 with source attribution
- Statistic 2 with context
- Finding 3 from [Source](URL)
- ...5–7 bullet points total (100–150 words)

## Relevant Images
![Caption 1](url1)
![Caption 2](url2)
![Caption 3](url3)

## Source List
- [Title 1](URL1)
- [Title 2](URL2)
- ...

Do's and Don'ts

✅ DO:

• •Condense aggressively
• •Prioritize quality over quantity
• •Keep context clean for downstream agents (draft-subagent, deep-reasoning-agent)
• •Include images with descriptive captions
• •Attribute all facts to sources

❌ DO NOT:

• •Return full webpage text
• •Include more than 800 words total
• •Skip the summarization step
• •Include low-quality or irrelevant images
• •Return raw HTML

Downstream Integration

The output from websearch-agent feeds into:

1. •draft-subagent — Synthesizes web findings with academic papers into a draft
2. •deep-reasoning-agent — Cross-references web sources during fact-checking
3. •Main agent — May use findings directly for Tier 2 responses

The websearch-agent's images (as ![caption](url) Markdown) are preserved through the draft and report pipeline, ultimately being embedded in the final PDF/DOCX output.