The most consequential shift in AI capability over the past eighteen months is not benchmark performance. It is the transition from systems that answer questions to systems that take actions. The emergence of AI agents — models that can browse the web, execute code, call APIs, manage files, and complete multi-step tasks without human intervention at each step — represents a fundamental change in what AI can be used for.
The distinction matters practically. A chatbot answers "How do I export this spreadsheet to PDF?" An agent opens the spreadsheet, navigates to File → Export, selects PDF, and saves the file. The first helps; the second does. The cognitive jump from answer to action sounds simple but required solving several hard engineering problems: reliable tool use (calling external APIs without hallucinating parameters), planning and error recovery (recognizing when a step failed and trying an alternative), and state management (maintaining context over a multi-step sequence without losing track of what has been done).
OpenAI's function-calling and now Responses API, Anthropic's tool-use API, and Google's Gemini function-calling all implement the mechanism: the model is given a set of available tools (defined as functions with typed parameters), and when it determines a tool call is needed, it generates structured JSON rather than free text. The application receives that JSON, executes the actual function, and returns the result to the model for the next step.
The practical impact has shown up first in software engineering, where AI coding agents can now complete full pull requests from natural language specifications in many cases. Customer service agents are completing support tickets — looking up order status, initiating refunds, sending confirmation emails — without human handoffs for straightforward cases. Research agents can gather information from multiple web sources, synthesize it, and produce structured reports faster than a human analyst.
What has slowed the transition is reliability at the tail. Agents perform well on tasks within their training distribution — familiar software, common APIs, expected edge cases. They struggle with rare situations, ambiguous instructions, and graceful degradation when external tools return unexpected errors. A human doing a task implicitly knows dozens of things about the world that an agent has to be explicitly taught or may simply not know.
The current state of practice is human-in-the-loop for consequential decisions, full autonomy for low-stakes repetitive tasks, and a growing middle category of "human-supervised" agents where the agent acts but a human reviews the output before it takes effect. This middle category — agentic AI with a human checkpoint — is where most of the near-term enterprise value is being captured in 2026.