When Agentic AI Comes Calling

I recently set out to answer a question that had been nagging at me: if an AI agent showed up at Webdancers’ front door, could it actually get anything done? Not “can it read the page” – I know it can do that. But could it look up what we offer, understand whether we’re a good fit for a project, and hand off a qualified inquiry? I didn’t know, and the only way to find out was to build the thing and see.

What is Agentic AI?

As a fairly new buzzword, it’s not well understood, so let’s define our terms. A regular AI chatbot answers questions. An agentic AI (i.e. an AI agent) takes actions on your behalf. It doesn’t just tell you what a form requires, it fills it out; it doesn’t just summarize a company’s services, it can go look them up on the company’s own systems and then act on what it finds. “Agentic” means the AI has access to tools it can actually use, not just talk about. That distinction turns out to matter a lot, as I discovered.

What I built

I put together a small prototype using something called MCP (Model Context Protocol), a standard that lets AI tools like Claude or Gemini connect to real data and take real actions instead of just guessing. I gave it two capabilities. The first could look up Webdancers’ services, portfolio, and client endorsements. The other was more active: submit a project inquiry on a prospect’s behalf, actually sending me an email. I purposly built it using my own business, for a couple of reasons. First, I have easy access to all the data and second, if I’m going to offer AI Search Readiness to clients, I ought to be willing to let an AI agent poke at my own site first.

For testing, I used Claude Desktop, the app version of Claude that you can install on your own computer. I configured it to point to a custom built MCP server, also running on my computer. When questions were asked on Claude Desktop, the MCP server reached out to the Webdancers site to retrieve the answers.

With that set up, Claude Desktop knows that my two new capabilities exist. I could then talk to it like I would any AI assistant: “I’m a nonprofit looking for a website redesign, can you tell me about Webdancers experience with nonprofits?” or “Go ahead and submit an inquiry for this project.” All interactions take place in a regular chat window, with the agent quietly reaching out to my WordPress site in the background whenever it needs real information.

This should not be a surprise: AI agents make mistakes

This one really got my attention. During testing, the AI agent submitted the same project inquiry twice. Not because I asked it to but because the agent itself decided, on its own, that its first attempt might have failed, and decided to try again. I ended up with two duplicate emails from one test inquiry, and a very clear lesson: an agent that can take real-world actions can also second-guess itself into doing the wrong thing. I fixed it with an explicit instruction telling the tool not to retry on its own, but that’s a patch, not a permanent solution. The proper fix is building duplicate detection into the system itself, which is on my list for later.

I also spent an embarrassing amount of time chasing what looked like a caching bug in the way endorsements were displayed. I’d update the content, check the output, and it wouldn’t change. I checked the CDN. I checked Redis. I checked for duplicate plugin files. The actual cause was a small text-cleanup function I’d written myself, which was producing identical-looking output even after a real edit. The lesson: when something stops updating right after you change code, check the code before blaming the infrastructure.

What this means in the real world today

There’s currently a chicken and egg problem with an autonomous AI agent discovering the capabilities of an MCP equipped website. I spent considerable time exploring how an AI agent might discover a website’s capabilities “in the wild” and the answer, as of mid-2026, is that it mostly can’t. The standards for AI agents to find and use these tools on small business sites are still very immature; one relevant specification published just last month has seen almost no real-world adoption. So if someone tries to sell you on “AI agent readiness” as an urgent must-have this year, be very skeptical. What does matter right now is making sure your site is legible to AI systems in the ordinary way: clear structure, accurate information, and content that answers the questions a site visitor — human or AI — would actually ask. That part isn’t speculative at all. It’s just good practice, and it’s available to you today.

AI Disclosure: I use Claude (Anthropic) to synthesize notes and conversation threads into a working draft, which I then rewrite, expand, and edit before publication. All content reflects my own views and judgment.