Tony Stark with Iron man suit — marvel.com

I Didn't Care About AI — Until It Got Agentic

Jul 4, 2025

When GenAI started gaining traction, there wasn’t much pull for me to dive in. The results were fascinating — it showed how far we’d come — but that same part of AI also led to unpredictability. While that unpredictability is still there, it seems like we’re beginning to overcome it. That’s what led me to agentic AI.

The Shift from GenAI to Agentic AI

GenAI could generate content—but that was often where it stopped. Agentic AI brings a shift: it can follow through. It doesn’t just respond, it gets things done. For example:

Deterministic: Agents aren’t guessing. They’re wired to follow clear steps, make calls, use tools, and actually complete tasks with some predictability.
Awareness-Driven: Whether it’s knowing what’s already happened or what tools are available, agents come with context. That lets them act a little more intentionally—not just repeat prompts.
Reliable Enough to Trust: You can hand over repetitive stuff—support flows, report generation, checks—and the agent won’t drop the ball every time. That alone is a big shift.

Challenges and Opportunities

Agentic AI is full of promise, but still early in its adoption. The path forward depends on solving a few hard problems—and seizing what’s clearly becoming possible.

Challenges

Limited Accessibility: While some companies have built production-ready agentic systems, they remain niche, expensive, and often tightly coupled to internal infrastructure. For smaller teams, this raises questions of cost, trust, and long-term reliability.
Brittle Architecture: An agent’s success depends on the right mix of LLM, tools, and context. If any one layer misfires—poor tool APIs, missing memory, bad outputs—the whole system can break down.
Fragmented Ecosystem: Most frameworks are Python-first, which limits accessibility for frontend or web developers. This slows down creative adoption in web-first teams.

Opportunities

Agentic SaaS: Software can now operate more like a teammate—proactively helping users, automating browser tasks, or interacting across platforms. This unlocks entirely new product categories.
10x Output in Workflows: Agentic tools are already transforming software dev—many IDEs now generate, refactor, and review code automatically. Similar impact is brewing in support, ops, and more.
Push for Purposeful Models: Unlike GenAI, agentic systems benefit from smaller, faster models—ones that are tuned for utility, not just fluency. This is driving innovation in the open-source and fine-tuning space.

What Building With Agentic AI Taught Me

Since March 2025, I’ve been exploring agentic systems—mostly through side-projects and prototypes. Here are a two lessons that stuck with me:

Crafted Intents Matter: Agents don’t thrive in open-ended chaos. You need a defined scope, a clear role, and (if using multiple agents) a deliberate coordination pattern—like Supervisor, Swarm approach, or Pipeline. Without that, you’re just paying for loops. Early experiments worked, but costs ballooned fast. Building lean PoCs with baseline agents helped keep things grounded.
The GenAI–Agentic Loop: GenAI often shines in generating ideas, drafts, or suggestions—surfacing gaps in workflows or content. Agentic AI complements this by taking action: automating fixes, following up, or completing tasks. Together, they form a loop where one identifies opportunities, and the other executes on them—making the system feel more useful and durable over time.

Conclusion

Agentic AI isn’t the future—it’s already here, just not evenly accessible yet. The pieces are in place, but it’s still early: expensive models, patchy tooling, and a steep curve for smaller teams.

But the direction feels right. It’s not just about generating things—it’s about doing them. And once the barriers drop—lighter models, better frameworks, clearer patterns—more folks will start building with it. That’s when it really gets interesting.