Agentic Scouting

About this episode

We are Nick and Jesse, two brothers running a marketing agency where AI agents do a real share of the work, and this show is what we actually use and what did not work.

Nick built an agent that reads every new business filed in Connecticut in a week, 1,117 of them, and hands back the 130 that are close to Hartford, have no website yet, and are worth a phone call. He runs it live for Jesse and explains each step, so you can point the same pattern at your own customers, your competitors, or anything else you keep Googling by hand.

Nick calls this agentic scouting. A search happens once. A scout keeps looking. He breaks it into six steps: search, filter, check, judge, decide, and remember. Search pulls the data from a source. Filter removes anything that breaks a rule you set up front, like being more than 20 miles away. Check looks up facts about what is left, such as whether the business has a domain, a live website, or a Google Maps listing. Judge gives those facts and your brief to a model, here Claude Sonnet 5 reading ten businesses at a time, and gets back keep or cut with one sentence of why. Decide is a person making the final call. Remember saves every outcome so the next run skips what it has already seen.

The part most tutorials skip is picking the data. Nick planned this scout with Ada, his assistant built on Claude Code, and they looked at business filings from 26 states before choosing one. The demo says it plainly: 22 of them failed on price, freshness, or terms. Connecticut publishes its filings for free, refreshes them daily, and includes an email on every record.

Then the run. 1,117 filings come in. 263 are within 20 miles of Hartford. Seven checks run on each one. Sonnet 5 reads them in groups of ten, and 130 land on the call list with a reason next to each. Nick also shows the 987 that were cut and why, reads the full judge prompt on screen, and finds a counting bug in his own demo while recording. A second scout, started live at the top of the episode for a smaller town, finishes while they talk.

Jesse brings the agency side. They run dozens of these every week for client work, mostly to watch what competitors are doing, and he is clear about what it takes: you have to build every step for your own use case, and your first scout will be a bad one.

Links

Claude: https://claude.ai

Claude Code, which Ada is built on: https://www.anthropic.com/claude-code

Our episode on Ada, Nick's AI chief of staff: https://www.youtube.com/watch?v=lbfFbKJ0REA

Connecticut's open data portal, where the state publishes business filings: https://data.ct.gov

Our newsletter: https://notreplacedyet.ai/#newsletter

Everything else we make: https://nryet.ai

Timestamps

0:00 Cold open

0:28 Welcome and what this show is

0:57 Getting agents to do your Googling

1:31 Why scouting and not searching

2:03 The six steps: search, filter, check, judge, decide, remember

4:26 The judge step: Claude Sonnet 5 reads ten at a time

5:27 Decide and remember: a human in the loop, and memory

6:21 Is agentic scouting a real term

6:48 How this scout was planned with Ada and Claude Code

9:57 Why Connecticut: 26 states checked

12:32 How many scouts we run every week

13:57 The dashboard, and starting a live run

14:41 Replay: new businesses within 20 miles of Hartford

15:35 Search and filter: 1,117 filings down to 263

17:28 How careful to be depends on the stakes

18:20 Seven checks on every business

20:34 The judge prompt, and a bug found live

22:32 The call list: 130 worth a phone call

23:09 Every business we cut, and why

23:47 Build every step for your own use case

26:03 The live run comes back

27:48 Where scouting goes next

29:42 Schedules and watching competitors

31:32 Who can use this at work and at home

32:43 Jesse: fifteen years of wanting this

33:25 Wrap: put these tools to work

This episode is sponsored by NonFiction Agency: https://nonfiction.agency