Understanding AI agents, automation, and MCP with Brad Klingenberg, Chief Algorithms Officer at Naro
The speed of change in AI is dizzying, and nowhere is that more apparent than in the rise of agents. As this technology explodes into the mainstream, leaders across industries are asking the same question: How can agents augment our teams and unlock new levels of productivity?
But amid the buzzwords and rapid product launches, it’s easy to lose track of what these technologies actually are and what all these terms stand for.
So we sat down with our Chief Algorithms Officer, Brad Klingenberg, to unpack the real story behind agents, MCP (Model Context Protocol), and automation. Brad has a rare gift for translating deeply technical ideas into clear, approachable explanations. Whether you’re a marketer, a builder, or just trying to keep up, this conversation will ground you in the fundamentals.
If you want to learn more about MCP and Agents for marketing, make sure to join our webinar on Thursday, August 7 at 11 am PT / 2 pm ET. Register here.
Q: Let’s start with the basics. What is a language model, and how does it relate to agents?
Brad:
A large language model is just a big neural network that takes in text and outputs text. Tools like ChatGPT, Claude, or Gemini are trained to take instructions, like “write me a blog post” or “summarize this transcript,” and generate helpful responses by using information encoded in the model and the prompt you provide.
That’s the core building block. Everything we’re talking about — automation, agents, tooling — starts from that basic ability.
Q: Everyone’s talking about agents now. What is an agent, and how is it different from just talking to ChatGPT?
Brad:
The experience most people have with language models is conversational. You type something, it responds. That’s useful, but limited. An agent builds on that by wrapping the model in a framework that adds two key things:
- Tools – A language model by itself can’t search the web or connect to a database. It just outputs text. But with an agent, you can give it tools. For example, if it outputs something like “WEB_SEARCH: Acme competitor”, a wrapper sees that, pauses the model, does the search, and returns the results back into the conversation.
- Reasoning and memory – An agent can talk to itself through a task. It doesn’t just follow a rigid script. It tries something, sees what came back, and decides what to do next. It’s self-directed.
Q: Can you give a concrete example of how this agentic reasoning works in practice?
Brad:
Sure. Let’s say you’re researching a competitor. An automation might take a keyword, run a search, and return the results. That’s it. But an agent might start with one query, see what came back, realize it’s too generic, and then run a second, more targeted query. It can reason through that flow.
It’s like how you’d use Google. Rarely do you search once and get the perfect result. Agents mimic that iterative behavior.
Q: So where do automations fit into all this?
Brad:
Automations are rule-based. They follow predefined steps. That can be great for reliability and speed, but they don’t adjust based on the output. Agents introduce flexibility. They reason. They adapt.
In our case at Naro, we have both. For example, our answer service is more of a structured automation. But we also have an answer agent that’s more open-ended, it will keep poking around until it’s satisfied with the result.
Q: What role does context play in all of this? What is “context engineering”?
Brad:
Context engineering is about getting the model the information it needs to succeed. Back in the early days, people talked about prompt engineering. Now it’s broader. It’s about what data you include, how you format it, what’s relevant, and how you retrieve and combine it.
If you ask the model to write about a product launch, you need to give it enough context: the feature, the positioning, the competitive landscape. Without that, you’ll get fluff.
Q: That leads us into MCP — Model Context Protocol. What is it, and why does it matter?
Brad:
MCP is an open protocol, started by Anthropic, that standardizes how language models call tools. It’s been widely adopted, even by competitors. Think of it like HTTP, but for models to interact with external services.
It makes it easy to plug in tools, think weather APIs, web search engines, or custom services like ours at Naro. You don’t need to rebuild the integration for each model. It democratizes access to tools, and it makes building agent workflows faster and more powerful.
Q: How does Retrieval-Augmented Generation (RAG) fit in with all of this?
Brad:
RAG is one approach to context engineering. It means retrieving relevant information — from documents, databases, etc. — and feeding it into the model’s context window.
If you have internal docs that the model wouldn’t have seen during training, RAG lets you find and insert the relevant snippets so the model can use them. It’s useful, but it’s just one type of tool. MCP lets you go beyond that, not just retrieval, but actions too, like sending an email or updating a CRM.
Q: Let’s bring it all together. What is the Naro Agent, and what makes it unique?
Brad:
The Naro Agent is an intelligent interface that enables users to interact with their calls and content using natural language and leveraging a suite of tools to analyze and draft, to name a few.
It combines call and content data, things like sales conversations, customer emails, internal decks, marketing materials, and external web research, and reasons across them. It can find insights, identify gaps, suggest messaging, and even generate content.
The real magic is in orchestration. The agent uses multiple internal tools (via MCP) and delegates work to specialized sub-agents, like a search agent or a summarization agent. Each one is expert at a specific task, so the primary agent can focus on solving the user’s broader goal.
Q: How is this better than just creating your own custom GPT in ChatGPT or Claude?
Brad:
Custom GPTs are great, but they’re limited. You can upload some documents, give it a few instructions, and it works… to a point. But you’re not indexing 10,000 sales calls. You’re not version-controlling your internal documents. You’re not verifying quotes or detecting hallucinations.
Naro is purpose-built to deeply understand your conversations and your content, and how they relate. It’s designed for marketers and sales teams who want trusted, up-to-date, actionable output at scale.
Q: Final thoughts?
Brad:
This space is moving fast. But at the end of the day, success with AI comes down to context — getting the right information to the model at the right time, in the right format. Agents, MCP, RAG, orchestration — these are just tools for making that happen more intelligently, more scalably, and more securely.
