Adoption numbers tell one story. The underlying technology tells another. A November 2025 report by MIT Sloan School of Management and Boston Consulting Group found that 35 percent of surveyed businesses had already deployed AI agents, while another 44 percent planned to implement agentic AI in the near future. That is a striking level of momentum for a category that most people, including many deploying it, struggle to define precisely. Understanding what agentic AI actually is, where it works well, and where it genuinely falls short is more useful than tracking the adoption curve alone.
From Generating Text to Taking Action
The clearest way to understand agentic AI is to contrast it with the generative AI tools most people already use. Systems like ChatGPT or Claude produce outputs: text, images, code, summaries. They respond. Agentic AI, by contrast, acts. It can book a flight, interact with a website, execute a sequence of steps in a software environment, or control a physical system like a robot. The distinction is not cosmetic. It changes the nature of the risk, the required oversight, and the kind of trust a user needs to extend to the system.
What most people do not realize is that the majority of AI agents on the market today are built on the same small set of foundation models. A company takes a model like Claude, wraps it with specific tools and memory capabilities, and deploys it for a particular use case. The “agent” is largely the wrapper: the tools it can access, the data it can remember, the actions it is permitted to take. This architectural reality, explained by Phillip Isola, an associate professor in MIT’s Department of Electrical Engineering and Computer Science and a member of CSAIL, helps explain why agentic AI has spread so quickly. The hard part, building a capable foundation model, was already done. The remaining work is integration and configuration.
Where It Works, and Where It Struggles
Coding agents represent the clearest success story in agentic AI to date. The reasons are instructive. Code is a domain where correctness can be verified automatically. An agent can attempt a solution, run it, check whether it works, and iterate. That feedback loop is what makes learning tractable. When the environment provides clear signals about success and failure, agents can improve through trial and error in a way that is difficult to replicate in messier, more ambiguous domains.
The training data problem is the central bottleneck everywhere else. Consider something as apparently simple as booking a flight online. The sequence of actions required, knowing where to click, how to handle errors, when to call and negotiate, is not well documented in any dataset. Agents attempting this kind of task must learn by exploring real environments, which are difficult to model and unpredictable. This is not a minor engineering challenge. It is a fundamental constraint on how quickly agentic AI can expand beyond the domains where it already performs well.
Isola draws an important distinction between agentic AI and analytical AI. Systems that model outcomes, surface patterns, or inform human decisions are not agents in the technical sense. They do not act autonomously. For high-stakes domains such as medicine, security, and high-level business policy, this kind of decision-support role may be more appropriate than full automation, both because the technology may not yet be ready and because the level of human accountability required in those contexts makes full delegation genuinely problematic.
The Risks That Get Underestimated
This is what most coverage of agentic AI misses: the risks are not only about what agents do wrong on their own. A significant share of the risk comes from how humans use them.
The ease of agentic tools creates a specific failure mode. When a coding agent can generate working software from a vague prompt, the temptation is to skip the verification step. Bugs get introduced. Private data gets exposed. Isola notes this is already happening. The agent may be performing exactly as designed; the problem is that the human did not define the task carefully enough, or did not check the output rigorously enough, because the tool made it feel like the hard work was done.
There is a longer-term concern layered on top of this. When people routinely delegate cognitive tasks, including coding, mathematical reasoning, and research, to automated systems, the skills required to perform those tasks independently may atrophy. The risk is not hypothetical. It is a known pattern in other domains where automation has reduced the frequency with which humans practice a skill. The concern is compounded by the fact that the technology is not yet reliable enough to fully replace human judgment in many of these areas. Delegating too early, before the system is ready, creates a situation where neither the human nor the agent is fully capable.
In Short
Agentic AI is not a new category of model. It is a new way of deploying existing models: giving them tools, memory, and the ability to act rather than just respond. It works best where feedback is clear and errors are recoverable, as in coding. It struggles where training data is sparse and environments are unpredictable. The adoption rate is high, but the maturity of the technology is uneven, and the most underappreciated risks are not technical failures but human ones: insufficient verification, vague instructions, and the gradual erosion of the skills that oversight requires. The open question for the field is whether the next generation of agents will be built on the same language model architecture with better tools, or whether genuinely new architectures will be needed to handle the physical and continuous world. That question does not yet have an answer.
Based on reporting from MIT News AI.