Hermes Agent amb /learn i /goal: com enriquir un CRM amb dades citades

Check this out. Five companies, zero contacts. I'm going to show you how to teach an AI agent to master a deep research API it's never even seen before. We're going to unlock some serious data enrichment potential and we're doing the whole thing with just one single prompt. Watch what just happened. It read a CSV. It called an API it learned 30 minutes ago. It researched five companies. It populated every field. It cited its sources. That's what this video is about. By the end, you'll know how to teach any Hermes agent any API once so it scales and then hand it real work that it drives to completion.

Hermes agent has two commands that change how you use it, {slash} learn and {slash} goal. Here's the short version. {Slash} learn teaches it something, an API, a workflow, a business rule, once. And it remembers it forever as a skill. {Slash} goal gives it a standing objective and it keeps working until that objective is actually met, not until you get tired of prompting it, until it's done. Today, I'm going to show you both on a real problem. We're going to teach Hermes the Parallel Task API, an API for web research that returns structured data with citations, and then hand it a CRM export with five blank company records and tell it to go enrich them, all five, without me. There are two reasons this works. First, Parallel's API is designed for exactly this. You give it a schema describing what you want and it researches the web and returns structured JSON with confidence scores and source citations per field. It's not a search engine. It's not a scraper. It's a research API. Second, Hermes can learn that API once via {slash} learn and then drive it autonomously via {slash} goal. The combination is what makes this more than a cool demo. It's a repeatable workflow you can run every time your CRM gets new leads. And at the end, I'll show you the one thing most tutorials skip. How to take this same setup and run it for a second CSV from a completely different industry without reteaching anything. The skill is permanent.

Before we dive in, Parallel is actually six APIs, not one. The task API is the star of today's video, but the skill Hermes learns covers the whole surface. Here's what else is in the toolbox. The entire suite is what makes this useful beyond today's demo. When Hermes learns Parallel via {slash} learn, it's not just learning the task API. It's learning search, extract, chat, everything. Future sessions, different use cases, same skill covers all of it. Today, we're going deep on task API for CRM enrichment, but by the end, the skill is permanent. Tomorrow, you could ask Hermes, "Search for recent funding announcements in the developer tool space." Or, "Extract the pricing page from all three of our competitors and build a comparison table." And it'll know how. One {slash} learn, six APIs, infinite use cases.

Let's start from zero. This Hermes agent has never heard of Parallel. It doesn't know what a task API is, and it has no skills loaded. Watch. This is Parallel's full API documentation in one file. All the endpoints, all the parameters, and all the schemas. I'm going to point {slash} learn at it. Type {slash} learn and provide the URL of the Parallel AI docs. I want a skill that knows how to use every Parallel API. Search, extract, task, and chat. Focus on the task API for structured research with JSON schemas, and include exact endpoint URLs, required headers, and the async polling pattern. The Hermes agent will go to the URL, learn the API, and a few minutes later, it will know exactly how to use it. There it is. The Parallel Web skill is created. It's permanent and reusable. In any session from now on, I can reference this skill and Hermes will know exactly how to call the Parallel APIs. Search, extract, task, chat. All of them. It has the exact endpoints, the required parameters, the async polling pattern, and even batch execution via task groups. All of it. Let's do a quick demonstration. Ask Hermes, "How do I create a task API run?" Hermes responds with the exact curl command, endpoint URL, and parameter explanation. All pulled directly from the skill. One {slash} learn command. That's all it took. Hermes now knows the entire Parallel API surface. Let me show you what that's actually worth.

Here is our scenario. This is a CRM export from a fictional B2B SaaS company called Pipeline, which sells developer tools. Their sales team just imported five new leads, but all they have are company names. No context, no qualification data, nothing. The sales team needs to know, "What does each company do? Who is the CEO? How big are they, and how much funding have they raised? Where are they headquartered, and what industry are they in?" Normally, that would require a junior salesperson to spend half a day on Google. But we are going to do it with one Hermes command. First, however, we need a schema. Company name, business descriptive, CEO name, headquarters, employee count range, founded year, latest funding total, primary industry. Eight fields. Each one changes how you sell to the company, yet each one is a Google rabbit hole when researched by hand. Now, every single one is something Parallel's task API can research and complete with citations, in a single call.

Here's the slash goal command. One instruction that provides an outcome, guardrails, and a definition of done. But watch what APIs it uses. This isn't just one call. slash goal Enrich every company in CRM-export.csv. For each company, one, use parallel search to find their official website domain and Crunchbase profile. two Use parallel extract to pull their about page and key data. three >> [clears throat] >> Use the parallel task API with this JSON schema to produce a structured profile. four Write all results back to the CSV with source citations. Use task groups for batch execution. Don't stop until all five companies have every field populated. Watch the order. It's not jumping straight to the task API. It's searching first, finding the actual domains, then extracting, pulling the real pages, then feeding that structured context into the task API for the final profile. Check it out. Here's the output it generated. To start with, we only had the company name and these two fields right here. But now, look at this. It enriched all eight fields automatically. Everything from the business description and CEO name to the industry and even the source citations. We can dive into the details here and you'll see how it filled all of this in on its own using the APIs it learned from Parallel Web APIs. Super cool. So, here's a new CSV with healthcare companies. Completely different industry, right? But watch this. It's the exact same skill and the same Hermes agent. I didn't have to reteach it a single thing. See, the skill doesn't actually care what industry these companies are in. It already knows the drill. Call the parallel task API, pull for the results, and write everything back to the CSV. The domain changes, sure, but the workflow stays exactly the same. And honestly, that's the big difference between a quick one-off demo and an actual system. I taught Hermes once, and now I can enrich any CRM export I want. Doesn't matter what the industry is, using that same original skill. It's basically powerful automation on autopilot.

I want to be straight with you about the tradeoffs. The parallel task API isn't free. While search and extract are cheap, the task API using the base processor costs more because it performs multi-source research with cross-referencing. For high-volume enrichment involving thousands of records, you should batch your requests smartly, and perhaps pre-filter which records actually need deep research versus those that only require quick lookups. Keep in mind that the skill Hermes learned from the learn command is only as good as the docs it was trained on. If Parallel updates their API with new endpoints or change parameters, you will need to rerun the learn command with the updated documentation. This isn't a set and forget process. Additionally, the self-evaluation step in the Hermes learning loop tends to rate its own work generously. Before you point this at a real CRM with 500 rows, spot-check the first few results manually. Just as I showed you with the citations. Trust, but verify.

Here's what you just saw. Step one, learn an API once. Point Hermes at the docs, and it writes a reusable skill. Step two, set up your scenario with a CSV, a schema, and a clear definition of done. Step three, goal. With just one command, the agent drives the entire enrichment process autonomously. I will link the parallel docs, the schema we used, and the learn and goal prompts in the description. You can copy them, swap in your own CSV, and run this today. What is the first API you would teach your Hermes agent? Let me know in the comments. Thanks for watching, and see you in the next video.