5 min readJust now

–

The tech world has been buzzing about AI agents for months now. CEOs proclaim 2025 as “the year of agents.” Startups raise billions promising autonomous systems that will revolutionize everything from customer service to software development. But here’s the truth that industry insiders are beginning to whisper: we’re not quite there yet.

What Even Is an AI Agent?

Before we dive deeper, let’s get clear on what we’re talking about. An AI agent isn’t just a chatbot that answers your questions. Think of it more like a digital intern or assistant something that can understand a goal, break it down into steps, use various tools to accomplish tasks, and learn from the experience.

For example, instead of you manually searching for flight prices, comparing options, checking your calendar, and then booking an AI agent would do all of that for you. It would remember you prefer aisle seats, know your budget constraints, and handle the entire process from start to finish.

Press enter or click to view image in full size

The Reality Check from the Experts

Andrej Karpathy, one of the co-founders of OpenAI and former AI director at Tesla, recently dropped a bombshell in a podcast interview. Despite all the excitement, he believes we’re looking at a “decade of agents,” not a “year of agents.” His assessment? Current AI agents “just don’t work” reliably enough.

Karpathy described much of today’s AI agent output as “slop” work that looks impressive at first glance but falls apart under scrutiny. The problem isn’t that these systems can’t do anything useful. It’s that they’re not reliable enough to trust with important, unsupervised work.

Meta’s Chief AI Scientist, Yann LeCun one of the “godfathers” of modern AI shares a similar perspective. He predicts that within three to five years, we’ll see a completely different paradigm of AI architectures emerge. The current generation of systems, he argues, will seem obsolete.

Why Aren’t Today’s Agents Good Enough?

Here’s where things get interesting. The fundamental issues are: these models lack core cognitive abilities, aren’t truly multimodal (meaning they struggle to work seamlessly across text, images, and real-world actions), have no reliable memory, and can’t consistently handle complex computer tasks.

Let me break that down in simpler terms.

The Memory Problem: Imagine hiring an assistant who forgets everything you told them yesterday. That’s essentially how today’s AI agents work. They don’t have continuous memory each conversation is mostly isolated from the last.

The Reasoning Gap: Current large language models are good at manipulating language but not at thinking. They excel at recognizing patterns in data but struggle with novel problems that require genuine reasoning.

The Real-World Understanding Issue: These systems lack a true understanding of the physical world. LeCun points out there’s a long way to go before AI can match humans in understanding how the real world actually works. A teenager can learn to drive a car in about 20 hours of practice. We still don’t have fully autonomous vehicles despite billions invested.

Where Agents Actually Work Today

Despite these limitations, AI agents aren’t completely useless. They’re showing real promise in specific, well defined scenarios.

Coding agents are particularly effective at boilerplate work repetitive code that follows common patterns found frequently on the internet. They can accelerate development by handling the tedious stuff, freeing human developers to focus on the creative, complex problems.

Companies like those building on platforms such as ClaudeCode, LyneCode, Cursor are already seeing this in practice. By leveraging AI agents for specific, bounded tasks like generating API integrations or writing test suites and similar forward thinking companies are achieving meaningful productivity gains. The key is knowing where agents excel and where they need human oversight.

In healthcare, AI agents are being deployed as medical scribes to draft clinical notes from patient conversations. Airlines are using them to handle complex booking changes, coordinating flight availability and fare rules. These are all examples of agents working within clear guardrails, with humans in the loop to verify critical decisions.

The Path Forward

So if we’re not ready for fully autonomous agents, what’s next?

Researchers at Meta and other institutions are working on developing AI systems that build mental models of the world. These future systems might develop some level of common sense and learn how the world works by observing and interacting with it.

This represents a fundamental shift. Instead of just training AI on text from the internet, the next generation will need to learn through experience more like how humans develop understanding.

LeCun predicts the coming years could be the “decade of robotics,” where advances in AI and robotics combine to unlock a new class of intelligent applications. This makes sense. To truly understand the world, AI needs to interact with it physically, not just process text descriptions of it.

What This Means for Businesses and Developers

Here’s the practical takeaway: AI agents are a powerful tool right now for specific applications, but they’re not a magic wand that will replace your entire workforce in 2026.

Smart companies are taking a measured approach. They’re identifying tasks where agents can add value today data processing, customer support for common queries, code generation for standard patterns while keeping humans involved in oversight and decision making.

The developers building agent powered applications understand this nuance. Whether you’re working at an established tech company or building something new, the winning strategy is to design systems where AI and humans complement each other’s strengths.

The Bottom Line

We’re in an awkward middle phase of AI agent development. The technology is good enough to be genuinely useful but not reliable enough to be fully autonomous. As Karpathy put it, “We’re at this intermediate stage. The models are amazing. They still need a lot of work.”

The hype cycle has gotten ahead of reality, but that doesn’t mean the vision is wrong just that the timeline is longer than many predicted. By 2026, we’ll likely see more sophisticated agents handling increasingly complex tasks, but they’ll still need that human touch.

For anyone building with or investing in AI agents: focus on the practical, measurable improvements they can deliver today. Find the right problems to solve, build in proper oversight, and prepare for a gradual evolution rather than an overnight revolution.

The future of AI agents is bright. It’s just not quite as bright or as immediate as the headlines suggest. And honestly? That’s probably a good thing. It gives us time to figure out how to build these systems responsibly, understand their limitations, and create frameworks that make them genuinely useful rather than just impressively hyped.

The decade of agents has begun. Let’s build it thoughtfully.