Company Research

Tool Restriction (Critical)

ONLY use web_search_advanced_exa. Do NOT use web_search_exa or any other Exa tools.

Token Isolation (Critical)

Never run Exa searches in main context. Always spawn Task agents:

• •Agent runs Exa search internally
• •Agent processes results using LLM intelligence
• •Agent returns only distilled output (compact JSON or brief markdown)
• •Main context stays clean regardless of search volume

Dynamic Tuning

No hardcoded numResults. Tune to user intent:

• •User says "a few" → 10-20
• •User says "comprehensive" → 50-100
• •User specifies number → match it
• •Ambiguous? Ask: "How many companies would you like?"

Query Variation

Exa returns different results for different phrasings. For coverage:

• •Generate 2-3 query variations
• •Run in parallel
• •Merge and deduplicate

Categories

Use appropriate Exa category depending on what you need:

• •company → homepages, rich metadata (headcount, location, funding, revenue)
• •news → press coverage, announcements
• •tweet → social presence, public commentary
• •people → LinkedIn profiles (public data)
• •No category (type: "auto") → general web results, deep dives, broader context

Start with category: "company" for discovery, then use other categories or no category with livecrawl: "fallback" for deeper research.

Category-Specific Filter Restrictions

When using category: "company", these parameters cause 400 errors:

• •includeDomains / excludeDomains
• •startPublishedDate / endPublishedDate
• •startCrawlDate / endCrawlDate

When searching without a category (or with news), domain and date filters work fine.

Universal restriction: includeText and excludeText only support single-item arrays. Multi-item arrays cause 400 errors across all categories.

LinkedIn

Public LinkedIn via Exa: category: "people", no other filters. Auth-required LinkedIn → use Claude in Chrome browser fallback.

Browser Fallback

Auto-fallback to Claude in Chrome when:

• •Exa returns insufficient results
• •Content is auth-gated
• •Dynamic pages need JavaScript

Models

• •haiku: fast extraction (listing, discovery)
• •opus: synthesis, analysis, browser automation

Examples

Discovery: find companies in a space

web_search_advanced_exa {
  "query": "AI infrastructure startups San Francisco",
  "category": "company",
  "numResults": 20,
  "type": "auto"
}

Deep dive: research a specific company

web_search_advanced_exa {
  "query": "Anthropic funding rounds valuation 2024",
  "type": "deep",
  "livecrawl": "fallback",
  "numResults": 10,
  "includeDomains": ["techcrunch.com", "crunchbase.com", "bloomberg.com"]
}

News coverage

web_search_advanced_exa {
  "query": "Anthropic AI safety",
  "category": "news",
  "numResults": 15,
  "startPublishedDate": "2024-01-01"
}

LinkedIn profiles

web_search_advanced_exa {
  "query": "VP Engineering AI infrastructure",
  "category": "people",
  "numResults": 20
}

Output Format

Return:

1. •Results (structured list; one company per row)
2. •Sources (URLs; 1-line relevance each)
3. •Notes (uncertainty/conflicts)