2025代理型AI优势：热潮、现实与未来展望报告（英）

ust two years ago, the term ‘AI agent’ barelyregistered outside research labs. Today, AI agentsare one of the most hyped topics in tech—and it’snot hard to understand why. The pitch is a powerfulone: AI agents can take care of time-consuming tasks so youdon’t have to. They can plan, make decisions, and use toolsto take actions and achieve goals on your behalf. Proponentssuggest they might do anything from organizing yourbusiness meetings to maintaining your software systems.In time, some claim, AI agents will take on tasks ofextraordinary length and complexity. “I think that peoplewill ask an agent to do something for them that would havetaken them a month,” said OpenAI’s CEO Sam Altman latelast year. “And they’ll finish in an hour.”J Global investment in AI agent-related startups has alsoduly exploded, with more than $2.8 billion flowing into thespace in the first half of 2025 alone. Ninety three percent ofIT leaders say they plan to implement AI agents before theend of 2026, and mentions of ‘agentic AI’—a term typicallyused to describe an AI system comprising agents—havesoared by 275 percent in company earnings calls. To“AI maximalists,” AI agents don’t merely represent anincremental advance in what software can do, but agenerational leap—one that stands to do much more thansimply improve personal and corporate productivity. Thebull case is that agentic AI will reimagine how organizationsoperate, rewire how economies function, and open up newpossibilities for how we live and work. “Agentic AI is a transformative technological advance thatwill drive step-change productivity improvement andinnovation across industries,” says Gene Reznik, ChiefStrategy Officer at Thoughtworks. “It will allow enterprisesand governments to reimagine their business processes andcommercial models, unlocking new sources of competitiveadvantage and differentiation.” Gene Reznik, Chief Strategy Officer,Thoughtworks For business leaders, however, there is a challenge: For allthe promise, there is a counter-narrative to contend with.Skeptics say the reality doesn’t match the hype. A keycriticism is that agents are unreliable. Since they are builton language models (LMs), which have a meaningful errorrate, they can make mistakes, fabricate information, getstuck in feedback loops, and diverge from intent. Combinedwith their limited memory, agents struggle particularly intasks that have multiple steps. In one now notorious instance,an AI coding agent deleted a software company’s liveproduction database. Detractors also point to securityvulnerabilities that let hackers hijack agents to steal data,manipulate systems, and spread malware. Gartner believesover 40 percent of agentic AI projects will be canceled bythe end of 2027, “due to escalating costs, unclear businessvalue, or inadequate risk controls.” So what is going on here—are agents the future ofenterprise, or is it all just hot air? The truth is not so binary.Enterprises are already deriving value from AI agents, butthese are not the general-purpose, autonomous “digitalco-workers” that the industry rhetoric sometimes implies.Successful deployments tend to involve more constrainedsystems, human supervision, and specific, contained usecases: A practical approach that aims to mitigate currentlimitations. Realizing the broader transformational promisewill require overcoming a number of technical hurdles. Business leaders should seriously explore agents as theystand today, taking a clear-eyed view of what they are andwhere they can offer value across the enterprise. At the sametime, they should track developments and be prepared toact fast if the field makes further advances and the dreamof autonomous super-assistants starts to materialize. The reality ofagents today So where are we now with agents: What can they do, howare they being deployed, and where are there proven examplesof their effectiveness? To unpick that, we first need to clear up some nuancesaround language. People use the term agent to mean a wide variety ofdifferent things. At one end of the spectrum, some businesseshave rebranded traditional software like RPA as agents. Atthe other end, evangelists conjure visions of all-knowing,do-anything systems. Both can confuse matters. The formeris just “agent washing” to attract investment and PR, andthe latter is an ambition for the future rather than a reflectionof what’s possible now. In practice, businesses today are indeed using agents thatrepresent a genuinely new class of software: AI that doesn’tjust answer questions but can use tools and take actions toachieve a predetermined goal. It’s just that aside from somehighly specialized systems, these agents tend to betask-specific, operate within tight bounds and almost alwayshave a human in the loop. Look at what most companiesmean when they say they’re running an agent, and typicallyit more accurately resembles an LM-based workflow. That’s less grandiose than the rhetoric—but that cannonetheless be powerful. For