Multi-Agent AI Backend: How I Built BPOAI's Nine-Agent Server

An AI agent server kept apart from the main site

BPOAI is a directory and intelligence hub for the BPO industry that I led at AIDEVGEN. The public Next.js site is only half the story. Behind it runs a separate agent server: a small Express app whose whole job is to make the directory feel alive.

The main site never runs an agent inline. It just fires a webhook. A small helper POSTs to the agent server with a shared-secret header and swallows every error, so an agent outage can never break the main site.

POST ${AGENT_SERVER}/webhook/{event}
x-agent-secret: <secret>

The Express app exposes routes such as /webhook/search, /webhook/claim-submit, /webhook/profile-update and a few /webhook/cron/* endpoints. That clean contract lets the agents evolve, crash and redeploy independently of the customer-facing site.

Scraping leads with Apify when a search finds nothing

The most interesting flow starts with a failed search. Agent 1 (Search) logs every directory query to Supabase and Slack. If a search returns zero results, it sets a flag that fires Agent 2 (Scraper) in the background. The site gets its response immediately, and the scrape runs after.

Agent 2 is the workhorse:

1. Find companies Apify's compass/crawler-google-places actor finds matching BPO companies on Google Maps.
2. Read their websites The apify/website-content-crawler actor pulls the text from each company site.
3. Extract emails A regex-based extractor harvests contact emails, preferring role inboxes like info@ and sales@ on the matching domain.
4. Save the leads New leads are de-duplicated and written to Supabase.
5. Chain two more agents Agent 8 notifies anyone who requested that scan, and Agent 9 starts an outreach campaign for every lead that has an email.

An automated email drip that stops when someone replies

Agent 9 (Outreach) runs a three-email drip.

• A unique database index prevents the same lead being enrolled twice.
• An admin master switch can pause all outreach at once.

The clever part is the feedback loop. When a recipient replies, Resend's email.received webhook hits the server. Agent 8 stores the reply and calls back into Agent 9, which marks the campaign replied and stops the drip.

AI readiness scoring with Claude, and when a domain check is enough

Agent 3 (Scoring) fires on every profile update. It asks Claude to compute an AI Readiness Score from 0 to 100, based on the company's listed tools, services and profile completeness.

Agent 5 (Claims) handles ownership claims with a plain check: it compares the claimant's email domain to the company's website domain.

Weekly cron agents for AI news and analytics

Every Monday morning, two agents wake up.

• Agent 6 (News) scrapes BPO and AI headlines across a set of topics with Apify's Google search scraper and content crawler. Claude turns each article into an SEO title, a summary and a clean HTML body, saved as a pending draft for an admin to approve. It also writes one original "trending" article naming the week's top-scoring and newly added companies.
• Agent 7 (Analytics) compiles weekly stats, such as new companies, claims, top searches and highest scores, and posts a digest to Slack.

I wrote more about where to keep people in the loop in When not to let the AI finish the job.

Shared libraries that keep each AI agent small

Four small libraries keep the agents thin.

Each agent reads almost like a recipe, because the messy parts live here.

Making a multi-agent backend safe to leave running

Three decisions did the heavy lifting:

1. Fire and forget by default Webhooks respond instantly, and the real work runs in background tasks with .catch() guards. A slow scrape never times out a user request.
2. One driver per job Outreach follow-ups are driven only by the external cron, never by a second in-process timer, to avoid double-sending.
3. Humans gate anything public AI-written articles and ambiguous claims wait for approval.

Nine agents, one webhook contract and a handful of shared tools turn a static directory into something that grows, scores and follows up by itself. It is the part of BPOAI users never see, and the part I had the most fun building.