Agentic AI — What It Actually Is and Why I'm Building With It

I Built an AI That Watches Everything So I Don't Have To

I have an AI operations agent called Luca that monitors all my clients' live deployments around the clock. It watches for failures, checks error logs, cross-references Git commits, and sends me a Telegram summary when anything needs attention — along with a recommended fix. Most of the time, I just tap “approve” and it's handled.

My clients never know any of this is happening. They just know their systems stay up and issues get resolved fast. That's the point.

That's agentic AI. Not a chatbot that explains what went wrong. An agent that spots the issue, investigates it, and resolves it — before anyone else even notices.

The gap between “AI that talks” and “AI that does things” is enormous. ChatGPT made everyone comfortable with the first part. But the real shift is happening right now, and most people haven't noticed yet.

What Actually Makes AI “Agentic”

I hear the word “agentic” thrown around a lot, usually by people selling something. So here's what it actually means in practice. An agentic system has four things going on:

Autonomy. You don't give it step-by-step instructions. You give it a goal. “Monitor my deployments and tell me when something breaks” is a goal. The agent figures out how — which APIs to check, how often to poll, what counts as a failure, who to notify.

Tool use. This is the big one. An agentic system can reach out and interact with the real world. It can query databases, call APIs, send messages, read files, trigger deployments. Luca has seven tools — Supabase queries, GitHub, Vercel, Railway, alert management, Outlook email, and Telegram. It picks which ones to use based on what I ask.

Planning. Give it a complex request — “check all my client environments and tell me which ones have had errors in the last 24 hours” — and it breaks that into steps. Query the database. Loop through each client. Hit the Vercel API for each project. Aggregate the results. Report back. It works out the sequence on its own.

Memory. It remembers what happened in previous conversations. If I asked about a deployment issue yesterday and follow up today with “did that get resolved?”, it knows what I'm referring to. Context carries across sessions.

The difference isn't intelligence. Both use the same underlying model. The difference is that one of them can actually do things.

How I Actually Built This

I'm not a massive company with an AI team. I'm a solo consultant. So when I say you can build agentic systems, I mean you specifically — a small team or an individual with some technical chops.

Luca runs on Claude's Agent SDK. The architecture is straightforward: a set of tools defined as functions that the AI can call, connected to a Telegram bot for the interface. When I send a message, the agent reads it, decides which tools to use, executes them, and sends back the result. The whole thing runs on a simple cloud server.

The key bit is how the tools connect. There's an open standard called MCP (Model Context Protocol) that lets you plug services into an AI agent like USB devices into a laptop. I used it to wire up my database, GitHub, and hosting platforms so the agent can query them all through one consistent interface. If I want to add a new integration, I plug it in. No rewiring the whole system.

And here's where it gets interesting. Because those tools use a standard interface, I was able to take the exact same seven tools and plug them into Claude Desktop — the app I use day-to-day on my Mac. Now when I'm working in Claude Desktop, I can ask it to check a client's deploy status, pull invoice data, or query my database, and it uses the same tools Luca uses. No API costs, no separate service — it just runs locally. Same capabilities, different interface. That's the power of building on standards rather than one-off integrations.

What the Big Companies Are Doing (and Why It Matters Less Than You Think)

Yes, Salesforce has Agentforce. Microsoft has Copilot agents. Google is building agent frameworks into everything. These are impressive products and they'll handle a lot of enterprise use cases out of the box.

But the interesting story isn't what a $300 billion company can build. It's that the same underlying capabilities are now available to everyone. The models are the same. The APIs are the same. The protocols are open.

I built an agent with seven integrated tools, deployment monitoring, error tracking, and a natural language Telegram interface. Two years ago, that would have required a team of engineers and a six-figure budget. Today it's a solo project.

That's the real shift. Not “big companies are doing AI.” Big companies have been doing AI for years. The shift is that agentic capabilities are now accessible to businesses with five employees and a founder who's willing to invest a few weeks.

Multiple Agents, One Platform

This is already how I run my consultancy. My operations platform doesn't rely on a single agent doing everything — it has specialist agents that each handle a clear domain. A finance agent manages invoicing and cost tracking. A product agent handles technical queries about client projects. An admin agent takes care of emails and scheduling. A router decides who should handle each request, so the right agent picks up the right task automatically.

Think of it like a small team where everyone has a defined role. The result for my clients is that things just get handled — quickly, consistently, without me being the bottleneck. That's the real power of agentic AI: not one clever bot, but a coordinated system that runs like a well-staffed operation.

Where to Start If You're Thinking About This

Don't try to automate everything. That's the fastest path to a broken system and a lot of wasted time. Here's what I'd actually recommend:

Pick one multi-step process that annoys you. Something you do repeatedly that involves checking multiple systems, making a decision, and then taking an action. For me, it was deployment monitoring. For a client of mine who runs a hair salon, it was appointment follow-ups — checking who's booked, sending reminders, handling rescheduling requests. Find your version of that.

Start with human-in-the-loop. Don't give the agent full autonomy on day one. Have it do the research and preparation, then ask you before it acts. Luca asks me before rolling back deployments. My property tech client's agent drafts responses to tenant enquiries but waits for approval before sending. Build trust incrementally.

Use existing tools before building custom ones. Claude, ChatGPT, and others all have tool-use capabilities now. Zapier and n8n can handle a lot of the integration work without code. You might not need a custom agent at all — a well-configured workflow automation might do the job. Only go custom when off-the-shelf hits a wall.

Measure what matters. Track time saved, errors caught, decisions automated. Not “we're using AI” but “this agent handled 40 deployment checks last week that I would have done manually.” Concrete numbers keep you honest about whether it's actually working.

This Isn't Coming. It's Here.

I'm not writing about something that might happen in 2028. I use agentic AI every day to run my consultancy. It monitors my clients' infrastructure, investigates issues, manages my alerts, and handles routine communications. It's not perfect — it still gets things wrong sometimes, and there are tasks I'll never fully delegate. But it's already saving me hours every week, and the improvement in the last six months alone has been dramatic.

The question for your business isn't whether agentic AI is ready. It is. The question is which process you're going to hand over first.