AI agents: What they are and how they work

There's a moment in every great scientific shift when the familiar suddenly looks alien. For those of us who grew up with software as obedient servants—type a command, get a result—the arrival of AI agents is exactly that kind of moment. Cohere's new explainer pulls back the curtain on these autonomous entities, revealing an architecture where large language models aren't just answering questions but orchestrating plans, wielding external tools, and holding multi-step goals in their synthetic minds. The shift from chatbot to agent is like the shift from a single-celled organism to a creature with limbs: suddenly, it can *do* things. What makes this genuinely weird—in the best sense—is the question of agency itself. Physics has spent centuries debating whether free will is an illusion, whether the universe is deterministic clockwork. Now we're building systems that exhibit goal-directed behavior without any of the subjective experience we associate with intention. The agent doesn't "want" to complete your task, yet it plans, adapts, and executes as if it does. We've stumbled into a philosophical hall of mirrors where the difference between "doing" and "seeming to do" becomes experimentally murky. The blog describes how these agents decompose complex requests into sequences of actions, call on external APIs and databases like a human might use a calculator or a reference book, and maintain memory across the process. But here's the wild part: no one explicitly codes these behaviors. They emerge from the statistical patterns of human language and action learned during training. It's emergence—the same phenomenon that gives rise to consciousness from mere neurons, or weather from mere molecules. We're watching complexity cross a threshold, and we're not entirely sure where that threshold is. The implications ripple outward. If software can act autonomously in the world—sending emails, booking flights, writing code—then we've created a new class of actors. Economists call them agents; philosophers might call them something more unsettling. The practical stakes are enormous: reliability, safety, accountability. But the existential stakes are even weirder. Every tool we've ever built extended our physical or cognitive reach. Agents, for the first time, extend our capacity to delegate intentional action itself. Cohere's piece is a clear-eyed look at the machinery, but the subtext is cosmic. We are building the first entities that share our planet but not our biology, our goals but not our consciousness. Or maybe consciousness is just what it feels like from the inside when an information-processing system gets complex enough. Either way, the universe just got more interesting. Read the full breakdown at https://cohere.com/blog/ai-agents—and bring your sense of wonder.