AI Agents: From Answers to Execution
Artificial General Intelligence (AGI) has been a buzzword of 2024, with tech leaders making bold predictions about superintelligent AI. But while everyone's focused on when machines will match human intelligence, something more immediate and potentially transformative is emerging: AI Agents.
A simple online search of the term will render a ton of headlines from tech companies who are actively pushing out digital agents. We're inching closer to a world where AI isn't just responding to our queries, but actively engaging with us, making decisions, and taking actions on its own. AI’s newfound agency may just be its crucial steppingstone toward AGI.
What exactly is agentic AI, and why should we care?
Alright, let's break this down. AI agents are built on action models (as we wrote about earlier this year) that go beyond the question-and-answer capabilities of language models to complete discrete tasks in the real world. In other words, agents could be like digital butlers: they can book a flight based on your expressed travel preferences, order your groceries, or recommend the best fantasy football lineup. And they can do so without the need for highly specialized prompts, thanks to improvements in AI’s language processing capabilities.
Most AI systems today are designed to respond to specific commands or queries. They excel at processing inputs and delivering outputs based on predefined algorithms. Now, Agency? That's where things get interesting.
Agency is all about how much the AI can recognize your needs and act on its own. Without agency, it only does exactly what you tell it, but with agency, the AI notices your elevated stress levels, recognizes that tomorrow morning's internal meeting isn't crucial, and makes the independent decision to reschedule it. It sees you're exhausted and takes the initiative to order your morning ritual (bagel and Venti-whatever-seasonal-syrup-they’re-selling-these-days), timing the delivery for your mid-morning slump. That's agentic AI in action.
That’s not to say that all of this happens through one model. An efficient agentic AI system would actually be coordinating a series of autonomous bots to pull information, analyze it, manipulate apps, and make decisions. But, as many hope to see in upcoming AI releases on their smartphones, users can set all of that system in motion with a simple request, like “I need a new suit for a friend’s wedding on the beach.”
How should we think about this new tool?
As we enter the era of agentic AI, we’ll want the power to approve any and all decisions the bots make on our behalf. Beyond simple safety protocols, we need intentional planning that considers both the promises and perils of autonomous agents.
Consider that a highway engineer may attribute an accident to poor road conditions, while a police officer points to reckless driving—neither explanation contradicts the other, and both reveal different aspects of the same problem. Similarly, in AI development, we might program a medical diagnosis system with the primary goal of “protecting patient health,” but we cannot assume that an autonomous agent will always develop the right independent subgoals, such as deciding when to recommend additional tests or prioritizing certain treatments over others.
Responsible AI development requires both thoughtful initial goal-setting and continuous oversight of how digital agents interpret and pursue these goals in real-world situations. Before our bots can truly make the turn to bespoke butlers, they’ll require immense amounts of data, oversight, and corrections. Keeping a human-in-the-loop will be more than necessary.
Sci-Fi Thoughts... Are We Still The Main Characters?
When your personal AI knows you better than you know yourself, who's really in control? Picture this: you're at a bar with friends, each of you with your AI agent quietly chiming in through earbuds, suggesting the perfect witty comeback or reminding you that maybe that next drink isn't the best idea given tomorrow's schedule. Or, think about the book (or movie) I, Robot and what can happen when humanoid robotics (in development) is married with agentic AI (already here). We can only hope that we’ve learned something from that prescient story.
For now, we're just scratching the surface of a world where our digital copilots are no longer tools, but active participants in our social fabric. In a sense, it may not feel that new since we’ve always relied on others to help shape our decisions...and at least we still get to choose when to turn the bots off. For now.
Excuse me while I go convince my AI that one more cookie won't hurt…
