AI Agent Swarm (non-coding) in Action | Saves Business Thousands

Description

I wanted to show how to use Claude Code, subagents and Claude skills in a non-coding and business capacity. I used it to save a business from buying an AI solution that would have cost them tens of thousands of dollars a year. Actually wrote a post about this earlier but got requests to show so here it is. I was using Claude Code in Cursor to do batches of document extraction work on my local drive. One of our clients was wondering if they need to pay their current software providers to do some intelligent document processing vs. being able to do it themselves. I was looking at using Claude Code + LlamaParse to: - extract the specific information that the client needed - develop a confidence score on the extraction - rename the file based on the extraction - and then move the file into another folder where their software would pick up and automatically segment. The initial thought was to have one subagent do all the extraction and renaming. But then the first big aha moment is, "Why not create a sub-agent for each piece of information that needed to be extracted?" So, each sub-agent would automatically go to work concurrently on each document. So, it would be a total of 4 sub-agents extracting the specific piece of information with a confidence sub-agent to evaluate the score and a renamer sub-agent appropriately rename the docs based on the confident score. And the process worked quite well. But the second aha moment is, Why can't we batch this? Why can't we have sub-agents processing each document but instead of one at a time, why not 10, 20, 30 at a time For 10 documents, it would be 60 agents working at the same time with 6 agents working on each document each document. Once I found that worked, it got me thinking. For literally any repetitive task you do on your local drive, why not have a swarm of AI agents (or in this case, sub-agents) doing it concurrently and consecutively? And if you want to do this on any of your tech stack you can connect it to any MCP. Let's look at the time to do this for an average of 100 documents twice a month (which this client has to do) MANUALLY (One person extracting the information) - 2-5 minutes depending on the complexity of the document. - 6-16 hours a month USING AI ONE AT A TIME ( 2-3 minutes with a drag and drop 4 to 8 hours a month 6 SUBAGENTS ONE PDF AT A TIME (Minimal Human Oversight) 2-3 minutes 6-10 hours a month SWARM OF SUBAGENTS (Minimal Human Oversight) 1min per doc 3 hours per month Now I know tools like Lindy has their swarm to do this, but if you can do locally or your tech stack with an MCP and just pay token costs for many businesses, that alone is well worth it.