Company Research

Tool Restriction (Critical)

ONLY use web_search_advanced from 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:

• company → homepages, gargantuan amount of metadata such as headcount, location, funding, revenue
• news → press coverage
• tweet → social presence
• people → LinkedIn profiles (public data)

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