All work

AI system · 2026

Lead research agent

Finds and qualifies sales leads, with evidence for every claim. Drafts the outreach. Sends nothing.

Role
Design, build, deploy
Stack
Next.js, Supabase, Claude Agent SDK, Apify, Firecrawl
Status
Archived

Overview

A list of companies is cheap. The research before the first email is the expensive part, and so is sending something wrong under your own name.

Decisions

  1. There's no send tool.

    The agent can't send anything because it was never given a way to. Each stage also runs sealed off from the machine it's on.

  2. The bar stays where it is.

    When there aren't enough good leads, it searches wider. It doesn't lower the criteria, and if it's still short, it stops and asks.

  3. Spending waits for approval.

    Paid searches only start after the founder confirms the criteria. Every stage has a cap on steps and spend.

Architecture

Lead research agent · 9 steps
  • Person
  • Model
  • Tool
  • Check
  • Result
  1. Step 1, Human: Founder sets the goal

    Founder

  2. Step 2, AI: Criteria written

    Agent

  3. Step 3, Human: Founder confirms

    Founder

  4. Step 4, Step: Find companies

    Apify

  5. Step 5, Step: Read each site

    Firecrawl

  6. Step 6, AI: Qualify with evidence

    Sonnet 5

  7. Step 7, Logic: Enough leads?

    Check

    Then: step 4 (short? widen the search); step 8 (yes).

  8. Step 8, AI: Draft outreach

    Sonnet 5

  9. Step 9, Output: Saved, not sent

    In the app

Results

  • 12 / 10

    qualified leads against target

  • 48

    drafts, none sent

  • 72

    actions, all logged

From one full run.

Known limits

  1. Nothing checks timing.

    Only 3 of the 12 leads had a real reason to reach out now.

  2. Nothing reviews the criteria.

    Every later stage depends on what that step writes.