Build Your Own AI Agent Workshop with Dstillerys

(00:29) We're going to learn by doing. Melinda, Mark, welcome.

(00:48) Thanks for inviting us. I'm happy to see some familiar names and faces here.

(01:04) Since you shared thoughtful poll questions, I'll end the poll so we can see where things stand and you can reference that. Let me share the results. Curious to see what you have for us today.

(01:30) To the two of you who hadn't heard of Agentic AI but chose to join, thank you. These results are great. I'm here with my colleague Mark Jung. Our goal for this session is that by the end, you'll be able to answer yes to all of these questions. If you're still not there, feel free to reach out and we can help get you there. Let's get started.

(01:55) We are from Distillery, the AI targeting company. Distillery uses a multimodal AI approach to audience building. Our models learn across data modalities to build a coherent understanding of a brand's best customers and their digital behavior informed by web journeys, LLMs, CTV, search terms, and more. We work with inputs across those data types to build predictive behavioral models that can be activated as user segments, contextual curation, and bidding algorithms. We use multimodal AI to do targeting.

(02:25) Why am I here to teach you about building an AI agent today? It's important for more of us to build an intuitive feel for what AI agents can do. The best way to learn is to build one. At Distillery, we built an agent called DS1 that gives us one point of interface to capabilities like audience discovery, building, and activation. We've seen it save time internally. As we've opened up access externally, we're learning that Agentic AI does more than save time. It creates a way for companies to connect more deeply and seamlessly. That's what I want to share: how Agentic AI can strengthen and streamline connections between companies. You'll get a feel for that in this hands-on workshop.

(03:43) Plan: background on Agentic AI and how it changes how companies connect. Then we jump into the hands-on workshop. Mark will run that. Then we'll switch back to a concrete example and demo of DS1, Distillery's agentic interface.

(04:04) Important: did you get your login email? Check for the subject "Your unique login for today's Build Your Own AI Agent workshop." It has unique login info for the hands-on part. If you signed up late or can't find it, DM Mark now or put it in the chat. Mark will set you up. While I go through intro material, you have time to get settled.

(04:51) I'm going to talk about Agentic AI and what it is. What makes AI agentic? Agency. Two things: it pursues goals autonomously, and it takes actions. It can plan steps to accomplish a goal, and it can act on those plans, use tools, interact with external systems, and pull levers in the real world toward those goals.

(05:58) There are many types of agents. For marketing, two types are useful to think about. Knowledge agents autonomously complete informational tasks. They may have access to customized documents and instructions and output content. Action agents have access to tools that interact with other systems. These agents complete goals by pulling levers in the real world. This type is getting attention in general usage (e.g., booking tasks) and in digital advertising because they can actually do things. This is the kind of agent we'll make today: an agent with tools that can act.

(07:05) You've heard that Agentic AI saves time by automating tasks. I’m going to talk about connection. Agentic AI streamlines connections between companies. The ecosystem is complex with many partners. Agentic AI can make connections tighter, leading to faster and better results. I'll show three types of connections between an agency and Distillery enabled with Agentic AI.

(07:51) Distillery has an AI agent with AI tools to discover, build, and activate audiences. First connection: human-to-agent. In this example, the agency has direct access to a Distillery agent via a chatbot (e.g., Slack). The human can use the Distillery chatbot to discover, build, and activate audiences. This is the most common way platforms and agencies are using Distillery’s agentic AI because it’s quick to start and requires no agency tech. The agency gets instant access with smart guidance and instant iteration without back-and-forth emails. Value: faster and easier iteration for more customized results. We thought agentic was just automation, but we realized it tightens connection and improves results.

(09:23) Second connection: agent-to-agent. The agency has its own AI agent. When it needs Distillery tools, the agency’s agent reaches out to Distillery’s agent for that part of the task. For example, building a media plan and reaching out to Distillery’s agent to discover, build, and activate audiences.

(09:47) Third connection: MCP (Model Context Protocol). MCP is becoming part of the conversation for making use of Agentic AI in digital advertising. The agency’s AI agent connects directly to Distillery's AI tools. MCP makes tools available to be accessed directly by other agents. Computers already talk across companies via APIs, but APIs take weeks or months of custom code. With MCP, the agent understands how to connect, so you can set it up in minutes.

(11:12) Questions welcome. If you want to unmute or use chat, please do. While I send credentials, I’m sending a link and a [workshop+number@distillery.com](mailto:workshop+number@distillery.com) address with a password beneath it. You might need to copy the email and password by hand. Keep shooting messages.

(12:17) Question: using the word "agency" is confusing. Can you broaden MCP beyond advertising agencies? Answer: MCP is a way for an agent to connect to tools. If you have a custom agent at your company or you use Claude as an agent, you plug it into MCP. Another company exposes tools via MCP. You give the agent a URL where the tools live. It contains what’s there and how to use it. The agent figures out what tools exist, when to use them, and how. It’s like handing a developer an API guide, but the agent reads and integrates it.

(13:55) There may be a human driving an agent (e.g., you type into Claude). Claude decides when to call which tools, possibly across multiple MCP connections.

(14:16) Example: connect Claude to GitHub and AWS via their MCP servers, then access them quickly. Security matters. Use tools you trust and authentication. MCP servers require authentication.

(16:44) Question on agent-to-agent vs tool calling (MCP): think of agent A and agent B at different companies. Agent A reaches out to agent B, and agent B decides which tools to use, then returns results to agent A. It’s not both working together simultaneously; B executes, then returns.

(18:21) Question: difference between AI tools and AI agents? Tools are not making decisions or planning; they aren’t agentic. Tools perform functions (e.g., audience building). Agents decide when to reach for which tools and how to use them.

(19:41) Login status update: you’ll receive a sign-in link, a [distillery.com](http://distillery.com) email, and a password. Use that email to log in and the alphanumeric password. Reach out if you don’t have it.

(19:59) One more question: do you have one MCP endpoint or multiple? Distillery offers multiple endpoints depending on which toolbox we hand you. One entry point gives access to the tools in that toolbox.

(21:19) ADCCP: launched recently as a standardization for agentic AI in adtech. It doesn’t change connection patterns; it standardizes message formats for certain tasks across companies. It sits on top of agent-to-agent or MCP. Our example is a use case not covered by ADCCP. In the workshop, you’ll use an agent that connects via MCP to external tools so you can see how easy it is to set these up and give AI access.

(22:50) IT/infosec concerns: we usually start with human-to-agent via Slack, which narrows IT questions. Companies ready for MCP are deeper into Agentic AI and prepared for it.

(24:20) Workshop in n8n to visualize what’s happening when building and talking to an agent. OpenAI just launched an equivalent; Google has one too. Concepts translate.

(26:07) First, add an AI agent node. Then add the brain: the chat model (OpenAI) and memory. Memory enables back-and-forth and context. Default memory window works for most cases.

(31:59) You’ve basically rebuilt ChatGPT if you stop there. To go beyond Q&A, add tools. For MCP, add the MCP client tool and point it to Distillery’s MCP endpoint (https://mcp.distillery.com/mcp). Set HTTP streamable mode. Save the canvas.

(35:05) Now the agent can use the MCP tools. Ask: "Help me search for audiences related to artificial intelligence." The agent knows to call Distillery’s search tool and returns audiences.

(36:42) To assist further, add Gmail via MCP to email results to a teammate with activation paths. The agent composes and sends the email using the activation details from Distillery’s tools.

(38:33) We’ll send detailed instructions for any steps you couldn’t finish.

(38:52) Practical big picture: this shows how you can create an agent, customize it, ask it for what you need, and have it push results into your workflow (e.g., email) without copying and pasting. Now I’ll demo DS1, our full agent used internally and by clients.

(40:45) DS1 is a human-to-agent Slack interface to our discovery, building, and activation tools. Ask "What can you do?" and it documents itself. In this demo, it can search for audiences, do some custom building, and support activation.

(41:51) Audience discovery via first-party data: "Please help me find audiences that will perform for Large Clothing Retailer." The tool uses the brand’s first-party data to run a simulation across our catalog of 20,000 audiences, predicting lift versus a random baseline. Without running media, you see which audiences are predicted to perform.

(42:50) Results appear grouped. First-party audiences (retargeting), Distillery custom AI audiences modeled from first-party signals, other custom audiences based on valuable behaviors, and pre-built behavioral audiences from our catalog.

(43:37) To understand third-party audiences, ask for thematic groups with lift and highest-lift groups first. Talking to your data via the agent helps build the activation story and plan. Making this accessible through DS1 changes how partners use first-party data to find pre-built audiences to activate for higher performance.

(45:02) Custom audience building: "Help me find purchase-intent seeds for vitamins." It searches product page URLs on retail sites from panel data to find the best fit for purchase intent. Pick product URLs as seeds, then the agent kicks off a multimodal AI model to target user segments and inventory most likely to be interested.

(45:59) If needed, narrow: "Please narrow to women’s vitamins." Iterate instantly to refine the audience definition before modeling. Then: "Please build an audience using the top three seeds." The agent kicks off audience building with predictive multimodal modeling. The agent is an interface and action layer, not the modeling itself.

(48:24) Human skills still matter, but our goal is to make it hard to put garbage in by guiding inputs and constraining the tool to high-quality seeds and models. The agent shows what it’s doing and asks for confirmation.

(50:48) Activation example omitted for time. Key takeaway: these agents use tools and pull levers in the real world to accomplish goals. The agentic layer interfaces with powerful tools, helping users use them better and creating faster, tighter connections that lead to better results.

(51:27) Thanks for the great turnout and interactive session. We’ll share materials. Feel free to email Melinda and Mark to continue the conversation. We’re happy to return with updates.