CLOSER — Autonomous Revenue Agents

An intent-first AI SDR: a crew of AI agents finds buyers raising their hand on the open web, qualifies the real ones, and writes the first message - warm, not cold spam.

Comment

Inspiration

The hardest, most expensive part of any business is finding customers. Every "AI SDR" on the market — Artisan, 11x, Clay, Apollo — does it the same way: buy a contact list and blast cold spam. Expensive, contract-locked, and the prospect never asked for anything. But people describe their problems in public every single day — "how do I sync inventory across Shopify and Amazon?", "looking for a tool to automate invoices." That's a raised hand. We built the opposite of cold outbound: an AI SDR that catches warm intent the moment it appears.

What it does

You tell CLOSER what you sell and who your ideal customer is. ~15 seconds later you get a board of qualified leads — real people who just described a problem you solve — each with a buying-intent score, the reason it qualified, a research brief, and a personalized first outreach message ready to send. You watch four AI agents do it live.

How we built it

  • Four specialized agents, one pipeline: Prospector (turns your product + ICP into intent queries and scans the open web), Qualifier (scores fit + buying intent, drops sellers/spam), Researcher (briefs the pain), Outreach (writes the first message).
  • Claude Sonnet 4.6 via the Anthropic SDK powers the agents — structured JSON outputs for scoring + prompt caching on shared instructions.
  • Next.js 16 + React 19 + TypeScript; the pipeline streams to the browser over Server-Sent Events so you watch each agent light up in real time.
  • Live signal sources: HackerNews, Reddit, Stack Exchange. Deployed on Railway.

Challenges we ran into

  • Live data from a datacenter: Reddit rate-limits cloud IPs, so we added HackerNews and Stack Exchange as keyless, datacenter-safe sources, plus optional Reddit OAuth and a high-intent fallback so the board is never empty.
  • Quality over noise: an early version surfaced off-topic leads. We made the Qualifier strict (fit + intent thresholds), and the agents never tell you "don't contact" — if a lead reaches outreach, it's worth contacting.
  • Cost & abuse: it's a public, no-login endpoint, so we added per-IP and global rate limiting.

Accomplishments that we're proud of

  • A genuinely different wedge in a hot, validated market: intent-first (warm) instead of list-first (cold).
  • Real, watch-it-happen multi-agent transparency — you see every agent reason, not a black box.
  • It found real buyers in testing (a multi-channel seller asking exactly the question our example product answers) with outreach you could actually send.

What we learned

  • The ICP framing matters as much as the tech — buyers asking for help in public are gold; DIY builders aren't buyers.
  • A reliable public demo means designing for graceful degradation (fallbacks, empty states, rate limits), not just the happy path.
  • Streaming the agents' work turns a 15-second wait into the most compelling part of the product.

What's next for CLOSER — Autonomous Revenue Agents

  • CRM/Slack push, one-click send, and reply tracking.
  • More signal sources (LinkedIn, niche forums) and a saved-search daily digest.
  • Team workspaces and per-account tuning — turning the open web into an always-on warm pipeline.

Built With

Share this project:

Updates