Agentic AI: the new 'groundbreaking technology' of 2025

Summary

Chain-of-thought and retrieval-augmented generation had their moments in 2022 and 2023. 2024 was reasoning models and multi-agent collaboration. 2025 is now agentic AI, and we may be onto something with this one. We’re five months into the chaos, so it’s a good time to look at what stuck and whether we should care.

What agentic AI is, and why people are excited

Agentic AI is the first real attempt to make AI more than a glorified autocomplete. A chatbot answers your prompt and stops. An agent gets a goal and keeps making decisions until it gets there, changing course when things go wrong. It works on its own and cooperates with other agents in a world that’s messy and unpredictable.

That’s a pretty big upgrade. For decades, AI was brought up to be the world’s most diligent intern: something that can work faster than humans without tiring. Agentic AI is the real step towards that, except it turned out to be more like an intern that starts their own company and then tries to automate your job. Think of a chess engine that suggests moves, versus one that signs itself up for a tournament and books its own flights.

The closest historical parallel I guess I can give is with the early Internet. Every network used to be its own little island, and nothing talked to anything else. Then TCP/IP and HTTP came along, and suddenly everything was connected. Agentic AI is at that same point right now. It’s fragmented and chaotic, and it might be about to turn into something huge.

Example: the travel-planning circus

Say you want to plan a five-day trip from Beijing to New York. The old way, a single agent works through flights, hotels and weather one at a time, like a very polite but slightly dim assistant. With agent protocols, you get the whole circus. Each of those jobs gets its own agent, and the agents negotiate until they agree on an itinerary.

The protocols differ in how they run the circus. MCP keeps things centralized and tidy. A2A and ANP let agents make deals across company lines. Agora, the most ambitious, lets agents agree on a structured protocol from plain-language descriptions, so they can work together without one being designed in advance.

The payoff is a system that can adapt on the fly when your plans (inevitably) change.

Not all interns are created equal

To understand agentic AI, you have to get used to the idea that “agent” covers a wide range. We’re nowhere close to perfecting this tech. At the shallow end you have simple reflex agents, like Roombas and rule-based chatbots. At the deep end you have multi-agent systems, where groups of specialized agents negotiate and bicker their way through complex tasks.

Most agentic systems today are a Frankenstein’s monster bolted together from the same parts. They take in data from whatever sources they can reach. They keep memory, both short-term context and long-term knowledge, sometimes with the recall of an elephant and sometimes with the recall of a goldfish. They break big, hairy tasks into small steps. They call tools like APIs and databases. They take real actions instead of just talking. And some of them learn over time, occasionally in ways their creators didn’t expect.

Like most of machine learning, a lot of this design copies God’s blueprints. Research on LLM memory borrows ideas like System 1 and System 2 thinking and theory of mind. Tool use exists because of human limits and specialization: we don’t do everything ourselves either. And I haven’t even mentioned the elephant in the room, which is that neural networks were inspired by brains in the first place.

Why agentic AI isn’t running the world (yet)

So why isn’t agentic AI running everything already? Because it’s hard. Really hard. Here are the main problems.

Standardization

Imagine everyone has their own agent, or their own team of agents. To be really useful, those agents need to talk to each other and work together. Right now they speak dialects so different that even Google Translate would give up. The lack of shared protocols is the bottleneck, and it’s the main thing keeping agentic AI from taking over the world (for now).

Security and privacy

The more freedom an agent has, the more you have to worry about what it does with your data. Authentication and access control are the bare minimum, not nice extras. Something something with great power comes great responsibility.

Reliability

An agent is only as good as its last meltdown. Remember the robot civil servant in Korea that drove itself down a flight of stairs after what looked like a rough day at work? In high-stakes settings, “works most of the time” isn’t good enough.

Evaluation

There’s no Consumer Reports for agentic AI (yet). Benchmarks are scattered, and everyone grades their own homework. Competitors end up calling each other out, and official reports get corrected and revised all the time.

Plugging in new tools

Plug-and-play is still a fantasy. Most integrations are brittle and manual, and setting them up is about as fun as assembling IKEA furniture with half the instructions missing. What you end up with is a mini helicopter running on three pig hooves on a train track. It might work for now, but it’s seconds away from falling apart.

Then there’s the human side, including the uneasy feeling that we might be building something we can’t fully control. We’ve all watched too many sci-fi movies, and that doesn’t help.

My take

What excites me most about agentic AI is the intentionality, much more than the buzzwords or the VC pitches. We’re finally building systems that chase goals and change plans mid-run, instead of just answering a prompt. That changes how we think about “intelligence” in machines. For the first time, it feels like we’re sketching something more than a tool. Something almost alive.

But I’m also wary.

People throw the word “agent” around a lot these days, usually without much rigor. Is a function-calling LLM with a planning loop and memory an agent? Technically, sure. So is a glorified macro if you squint hard enough. I think the harder problem is building agents that know where their competence ends. I want agents that can say “I don’t know” and hand the task to someone else. More capability is easy to sell. Knowing your limits is the frontier.

And honestly, a lot of what gets sold as agentic AI right now is old ideas with a new label: finite-state machines wrapped in Python with a “GPT” logo glued to the front. Cool demos, but that’s about it. People are closer to making movies than documentaries. If this field wants to survive the hype cycle, it needs fewer sizzle reels and more system audits, and more principled certification instead of vibes-based evaluation.

Still, I’m optimistic. Behind the noise are the researchers who are actually in the weeds, building new protocols and fixing stupid bugs at 2am. They’re the ones pulling the field forward, and most of them aren’t doing it for clout or paper count. They do it because they can see the outline of something real and strange.

and I want to be there when it shows up.