One of the most useful AI agents I've built this year is a networking agent, which I use at Kwanda to build relationships with funders.
How it works
It understands our value prop, goes out onto the web, finds aligned funders along with the key person inside the organisation, drafts a light-touch personalised email for each one, and drops them into my Gmail drafts. In 5 minutes I have 10 emails ready to send.
Useful, but I still had to remember to run it.
How I improved it
This week I made it always-on. It now listens to my meeting transcripts. If I get off a call where I was talking about a new funding thesis, or where I mentioned I want to find people doing interesting work on agriculture in Nigeria, the agent reads the transcript five minutes later and goes looking. By the time I sit back down at my desk, there are fresh drafts waiting.
It has been a great way to connect with people I'd never have thought to reach out to.
Here's how to build it
- Start with a short value prop doc. Who you are, what you're building, and what you're looking for in a relationship. This is what makes the emails feel like you wrote them, so be specific about the work, the stage, and the kind of person you want to reach.
- Plug in a research tool like Perplexity, Apollo or Exa. This is how the agent finds aligned people and digs a level deeper to find the right person inside the organisation.
- Define what "aligned" actually means to you, whether that's geography, stage, sector, or cheque size. Without these filters you'll get a long list of generically relevant names instead of ten that are worth emailing.
- Connect Gmail with draft-only access so it never sends, and you review everything before it leaves your inbox.
- Write the email template. Keep it short with no ask. One line on why you're reaching out, one on what you do, one opening the door, and let the agent fill in the specifics from its research.
- Test it by hand first. Run it with something like "find me five funders doing X", read the drafts, and tune the prompt until the emails sound like you.
- Now add the always-on layer. Hook it into your meeting transcripts ( Fathom, Granola, Fireflies all work). After the call, it reads what you actually talked about, pulls the topics, and runs its search off the back of that. Drafts land in your inbox five minutes later.
- Give it a memory. A Notion database or Google Sheet of everyone it's already contacted. The agent checks it before drafting so you don't email the same person twice.
What I've learned tuning it
The value prop doc does most of the work. Early on mine was too broad and the emails came back generic. Rewriting it to be specific about our thesis, the projects we fund, and the funders we're a natural fit for changed the quality of the drafts overnight.
Keep the email template restrained. No ask, no pitch, no calendar link on the first touch. Just a real reason you're reaching out and a door left open. Anything heavier and the agent ends up writing something that feels like a cold email, which it is, but shouldn't read like one.
The always-on layer is the real unlock. The manual version was good, but I'd forget to run it for weeks at a time. Tying it to meetings means the agent works when my thinking is freshest, not when I remember.
Tools
Claude Code for the agent. Perplexity and Apollo for research. Gmail API for drafts. Notion for meeting notes and contact memory.
If you set this up you'll stop missing the people you should have been in touch with months ago.