I Spent a Weekend Testing the "Agentic AI" Hype. Here's What I Found.

Originally posted on LinkedIn · March 20, 2026

"Agentic AI" has been coming up a lot lately, especially in finance conversations. I wanted to understand what it actually looks like in practice, so I spent a weekend testing it myself.

On a Saturday. With coffee. And a healthy amount of skepticism.

What I found was surprisingly accessible and, by Sunday evening, genuinely interesting.

The difference between a tool and an agent

Most people in finance have used ChatGPT. It's impressive, but reactive. It responds when prompted. That works for one-off questions. It breaks at scale.

An agent operates differently. It holds a goal, connects to live data, makes decisions within defined boundaries, and acts without being prompted.

The simplest analogy I landed on:

It's the difference between a consultant you call and a chief of staff who briefs you before you've had coffee.

For finance functions, where value is tied to timely and reliable intelligence, that distinction matters.

What I actually built

I used OpenClaw as the agent framework, with Claude as the underlying model, both accessible and low-cost to get started.

I configured an agent connected to public data sources like FRED, market data, and financial news.

• It runs on a schedule so a briefing arrives before market open.
• It filters based on defined criteria instead of surfacing everything.
• It runs fully sandboxed on a personal setup.

By Sunday evening, I had a working prototype.

What stood out wasn't the setup.

It was the thinking required.

Deciding what to monitor and what makes output useful was far more challenging than wiring things together. That's a finance problem, not an engineering one.

On governance

The real questions aren't about capability. They're about risk.

I kept everything read-only, fully sandboxed, and separate from any work environment.

The bigger realization: defining what the agent should do, what it should ignore, and where it should stop is the hard part.

That's not technical. That's strategic.

Reflections

The technology is further along than I expected.

But access and cost aren't the constraint. Clarity is.

Knowing what to ask, what matters, and how to frame the problem still matters more than knowing how to configure the tool.

I came away more curious than I started.

If you've been exploring this space as well, I'd be interested to compare notes.