Telegram as the control plane
Each campaign gets its own Telegram topic, so parallel runs stay separated in one interface. The agent reports what it found and what it proposes; approvals happen in the same thread, from a phone.
Agents that source businesses from Google Maps, qualify them by web presence, build a mock site on Vercel, and report back through Telegram.
Running as recurring scheduled tasks, with human approval before any outreach goes out
Prospecting is mostly triage. Find businesses, work out which ones have a problem worth solving, then produce something concrete enough to earn their attention. Done by hand it stalls at step one. Agents do the sourcing, the qualification, and the first draft, and I approve what goes out.
Agents can do the finding and the first draft, and you still approve everything before it reaches a customer.
Each campaign gets its own Telegram topic, so parallel runs stay separated in one interface. The agent reports what it found and what it proposes; approvals happen in the same thread, from a phone.
Businesses are pulled from Google Maps and classified by web presence: no site at all, a site-builder template, or a site that no longer works. Only the qualified ones move forward, with their contact details.
For a qualified prospect the pipeline builds a working demo site and deploys it to Vercel. The first thing they see is their own business rendered properly, not a description of what could be done.
The drafted message and the live demo link are handed off for outreach once I approve them. Nothing leaves the system unreviewed.
One campaign is driven by Claude Code, a second by another agent runtime on the same design, aimed at a different niche. They run as recurring scheduled tasks across a self-managed VPS, a mini PC, and a Mac M1.
Tell me what you are trying to automate. A call is usually enough to scope it.