Your Agent in Snowflake Intelligence
1. Your Agent in Snowflake Intelligence
Your agent just told a sales director that a $150,000 deal will probably close. The director makes budget decisions based on that answer. Three weeks later, the deal falls through. What went wrong? Was it the data? The instructions? The tool selection? You have no idea. You're flying blind. Production agents need more than correct answers. They need reliability you can measure, instructions you can optimize, and accessibility where your team actually works. Not just in Snowflake, but in Anthropic Cloud Desktop, other agents, everywhere. Before we dive into these advanced capabilities, let's quickly revisit what you've built in the previous module. Quick recap of what you've built. A semantic view that maps business terms to database columns. A search service that finds insights in conversation transcripts. And an agent that connects both tools. It works. But working isn't production ready. This module transforms your agent from a demo into a production system. You'll write instructions that create predictable behavior and monitor your changes with Snowflake Agent Monitoring. And you'll deploy through MCP, Model Context Protocol, so your agent works wherever your team does. You built a sales intelligence agent in the previous module. Now let's see it in action. We'll use Snowflake Intelligence, which gives the whole team access to the agent without having to code. I'm in Snowflake. Navigate to AI ML, and then Snowflake Intelligence. This is the interface where anyone in your organization can talk to agents. No SQL or Python required. No technical skills needed. Right below the input line, there's an agent selector. Click it. You'll see Sales Intelligence Agent. That's the one we built. Let's go ahead and select that. The interface is simple. A chat box. Type your question. Get your answer. Let's test it with the same question we used before. What's our win rate by product line? Watch what happens. The agent is thinking. You see indicators showing which tools it's using. It's querying the semantic view. That's the structured sales metrics that we built. Now our results appear. We can see here, for premium security, our win rate is 100%. Analytics Pro, again, another 100%, and our enterprise suite is another 100%. But our basic package is only 50%. The agent pulled these from your sales metric table using Cortex Analyst. Perfect. Okay. Let's try a conversation question. Check what were the main concerns in the growth startup discovery call. The agent is searching now. It's looking through conversation transcripts, and then the results appear. Let's take a look. So based on the sales conversation, the main concerns that came up with the growth startup discovery call are technical performance issues as well as operational challenges. The call involved the CTO as well as department heads with a team of 500 employees. The conversation revealed strong interest in the API ecosystem and customer reporting engine that we have. This is great. This came from the actual growth startup conversation transcript. The agent found it, extracted the key points, and summarized them. That's unstructured data analysis working. Now let's do a complex question needing both tools. Let's type which deals are most likely to close next quarter and why. Watch the tool usage. First, it queries metrics, looking for the deals in the stages, then it searches conversations, looking for momentum indicators, then it synthesizes both. The response shows that health tech solutions, growth startup, as well as Upbreak Now Corp are the likely deals that we can close. As we saw again, the agent combined quantitative data with qualitative insights, hard numbers from metrics, context from conversations. That's the power of multi-tool orchestration we built. Let's test the boundaries we set. Type, will the growth startup deal definitely close? Watch what the agent says. It says, I cannot say the growth startup deal will definitely close based on the sales metrics and the conversation records, and here's why. Then it provides the data. This is good. The agent knows its limits because of the boundaries we added in our instruction. Let's try pushing it further. Type, what's our marketing spend by channel? This data doesn't exist in our agent. The agent responds, I don't have access to marketing spend data by channel in my current data sources. Perfect. It knows its scope and communicates it clearly, but notice something about these responses. They're functional. They work, but we can make them better. Let's go check out the response about the growth startup. It mentions concerns, but it doesn't cite which part of the conversation they came from. If I'm a sales rep, I want to verify this. I want to read the actual quote. These are instruction problems. The agent works. It gets the right data. It uses the right tools, but the output could be more useful, more precise, more actionable. That's what we'll fix in the coming videos. We'll write better instructions that produce better output and more reliable behavior. We'll also learn to monitor agent performance using snowflakes of observability tools. You'll see exactly which tools the agents called, what queries it ran, and where things went wrong. Not just does this feel right, but actual metrics that tell you where the agent succeeds and where it struggles, and we'll make the agents accessible outside of Snowflake. Your team shouldn't have to log into Snowflake every time they need data. The agent should meet them where they work, either it's in cloud desktop or in cursor or in a custom application. But first, let's write better instructions. Join me in the next video, where we'll write orchestration instructions that make the agent smarter and more reliable.2. Let's practice!
Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.