Why I'm Skeptical of Agentic AI (Despite Using AI Every Day)

Catching up with a peer recently, the conversation landed on—half-joking, half-incredulous—how I of all people, someone building software and systems with AI every day, don’t believe in agentic AI.

“You seriously aren’t using agents yet?”

I get the surprise. I’ve built real production workflows using structured prompts, modular AI tools, and memory scaffolds. I use AI to code, test, document, enforce rules, manage themes, even generate PRDs. I’m not an AI skeptic.

But I am skeptical of the agentic abstraction layer—the idea that we can just hand off a goal to an autonomous AI agent and let it plan, reason, execute, and adapt on its own. Not because I don’t understand it. But because I’ve lived through the limits of every piece that goes into it.

The Core Problem

Even with my own prompting, strict memory systems, and tight schema enforcement—there is still drift.

Now layer in: a second LLM trying to prompt another LLM. Tool calls that may silently fail or hallucinate outputs. Plans with no execution safeguards. Scratchpad memory that rewrites itself mid-run.

You don’t reduce fragility. You multiply it.

You’re trusting an LLM to prompt better than you, plan more reliably than you, debug and self-correct faster than you—even though you already know how fragile and hallucination-prone the underlying models are.

Let’s be honest: if you’ve ever tried to build something real with these systems, you already know. Most agentic demos are just Rube Goldberg machines for calling GPT-4 in a loop with fancier error messages.

When Agency Actually Works

I’m not anti-agent. I’m anti-fantasy.

Real agent-like behavior shows up when the task is low-risk and high-volume, the workflows are observable, recoverable, and testable, and the agent isn’t the planner—I am.

Don’t outsource intelligence. Outsource tedium.

Build a sync monitor that watches match logs and retries failed requests. Build a test runner that confirms UI flows against specs. Build a memory coach that asks, “Want me to store this pattern?” That’s agency with constraint. Useful. Measurable. Bounded.

What I Actually Use

Declarative prompt stacks. Type-safe, memory-aware scaffolds. Modular AI toolchains with clearly defined roles. Human-in-the-loop plans with CI-level enforcement. Schema-first design with rollback paths.

AI systems with accountability.

When I say I don’t believe the hype around agentic AI, I’m not being a hater. I’m just someone who’s actually been burned by AI drift—and decided not to add another drift-prone actor into the system.

If you’ve built something real with agentic AI that holds up under real-world complexity, I’d genuinely want to see it. But until then, I’ll keep building AI systems with grip, not glitter.