What Is Agentic AI Today, and What Do We Want It to Be?

Summary

Agentic AI—artificial intelligence systems that take actions in the world rather than simply generating content—has moved from research labs into mainstream business deployment. According to a November 2025 report by MIT Sloan and Boston Consulting Group, 35 percent of surveyed businesses have already deployed AI agents, with another 44 percent planning implementation. Unlike generative AI models like ChatGPT or Claude that create text and images, agentic AI systems perform digital and physical tasks—from booking flights to solving coding problems—by combining foundation models with specialized tools and feedback mechanisms.

According to Phillip Isola, an associate professor at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), most agentic AI systems today are built on the same foundational architecture: a large language model (like Claude) augmented with specific tools and the ability to remember context. The challenge is not the underlying models but training data scarcity. Systems that learn to navigate websites, handle edge cases, and execute complex sequences require either massive labeled datasets or—more commonly—trial-and-error learning in dynamic environments that are difficult to simulate.

Coding agents represent the most successful application to date, leveraging feedback loops where the system attempts solutions and validates results. However, experts caution that not all tasks should be fully automated. High-stakes domains—medicine, security, major business decisions—require a balance between AI assistance and human oversight, suggesting that the technology's maturity varies significantly across use cases.

• Training data scarcity is the primary bottleneck; most systems learn via trial-and-error in real environments
• Coding agents are the most mature application; high-stakes domains require human oversight rather than full automation

Editorial Opinion

Agentic AI represents the next evolutionary step in AI deployment, moving from passive content generation to active problem-solving in real-world systems. The rapid adoption rate (79% of businesses engaged or planning) signals genuine business value, yet the technology remains immature outside narrow domains like coding. The critical design question—where to automate versus where to augment human decision-makers—will likely determine whether agentic AI fulfills its promise or creates new operational risks. Responsible deployment requires understanding that 'capability' and 'readiness' are not the same thing.