Research Agent for Market Mapping

What this agent does

Given a seed company name or market category, this agent returns a structured competitive landscape: who the players are, how they position themselves, what the common patterns are, and where the gaps appear to be.

A task that previously took 4–6 hours of manual research now takes 8–12 minutes.

Agent architecture

The agent runs as a multi-step chain with a supervisor model and tool-using subagents:

Step breakdown

Step 1: Player identification

The supervisor breaks the input into a search strategy. For a seed company, it first identifies the category, then finds adjacent players using Exa’s neural search. For a seed category, it searches directly.

The player list typically returns 12–20 companies. The agent filters these against basic criteria (active website, appears to be a real product, not just a review site) before proceeding.

Step 2: Positioning data collection

For each company on the list, the agent fetches:

• Homepage headline and subheadline
• Pricing page (if public)
• “About” or mission statement
• Any case studies or customer logos visible on the homepage

This is the most expensive step — one Exa call per company. The agent batches these in parallel to stay within rate limits.

Step 3: Pattern recognition

The supervisor processes all positioning data together and identifies:

• Common positioning axes (e.g., “ease of use vs. power,” “SMB vs. enterprise”)
• Shared language patterns (terms that appear in 5+ positioning statements)
• Pricing model clusters (free tier, per-seat, usage-based, etc.)
• Who is targeting whom (explicit or implied customer segments)

Step 4: Gap analysis

The final step asks: given this landscape, what positioning is noticeably absent?

This is the most model-dependent step and the least reliable. The agent flags this output explicitly as “interpretive” rather than “factual” — it is useful for hypothesis generation, not as ground truth.

Sample output (truncated)

Category: AI meeting assistants
Players identified: 14

Positioning clusters:
  • "Never take notes again" — 6/14 players
  • "Built for [specific team]" — 4/14 players  
  • "Privacy-first / on-premise" — 2/14 players

Common language:
  action items, follow-up, transcript, integrates with, searchable

Pricing clusters:
  Free tier: 9/14
  Per-seat ($12–$25/mo): 8/14
  Enterprise / contact sales: 11/14

Apparent gaps:
  • No player prominently positions on async-first (meetings you were not in)
  • No player targeting finance teams specifically (despite high meeting load)
  • "Search across all past meetings" underemphasized given technical capability

Reliability and limitations

The agent is reliable on the factual steps (player identification, positioning data) and useful-but-imperfect on the interpretive steps (gap analysis).

Common failure modes:

• Thin positioning data: Companies with a single-page site or a marketing-heavy homepage give the agent less to work with. Output quality drops.
• Fast-moving categories: In very active markets, 20% of the competitive landscape may have shifted since the agent last ran. This is a timestamp problem, not a model problem.
• Stealth or pre-launch players: Exa searches find public pages. Companies not yet marketing themselves are invisible.

The agent is best used as a starting point, not a final answer. It eliminates 80% of the research time and surfaces the right questions. A human analyst still needs to verify the most important claims.

Infrastructure notes

The agent is implemented in Python using the Anthropic SDK with tool use. The Exa integration uses their search_and_contents endpoint for combined search + scraping. Total API cost per run: approximately $0.40–$0.80 depending on category size.