商业，满足代理AI-对自主和代理解决方案的信心（英）

ForewordIn Capgemini’s AI Futures Lab, we study advances in AI and the implications that thoseadvances will have on the world. It’s a fascinating and awe-inspiring work where weare confronted every day by the amazing possibilities coming over the horizon, but thepotential implications of autonomous and agentic systems are unparalleled. The reasonthat we’re so excited about this is not particularly because of the technology behind itor any particular AI model that’s used, but because it fundamentally shiftsour relationship with technology.For the entirety of our history with computing, if we wanted to make a computer dosomething we needed to describe in great detail how to solve that problem, either byprogramming it ourselves or relying on experts who could. That was, by definition, anexclusive arrangement where only people who understood technology could get themost out of it. The era of autonomous and agentic AI presents us with a new vision ofthe future, one where users of technology can command technology to solve problemsthat they themselves have no idea how to solve. This is the version of AI that we werealways promised by science fiction, where anyone can harness the full power of AI.The reason agentic and autonomous systemsare transformative is because they allow usersof technology to shift from defining the solutionto simply stating their problem.Robert EngelsHead of AI Futures Labrobert.engels@capgemini.com2 IConfidence in autonomous and agentic systems Mark RobertsDeputy Head of AI Futures Labmark.roberts@capgemini.com ExecutivesummaryToday’s technology landscape is quicklychanging. Autonomous and agentic AIsystems, systems that can make decisionsand take action on their own, arebecoming increasingly important. Thesesystems aren’t just a small step forward;they represent a major shift in how peopleinteract with and experience technology.Autonomy is a game changer. By allowingautonomous and AI agents to take actionswe unlock amazing new opportunities, butthese do not come without risks.This white paper builds on our2024 edition ofUnleashing Confidencein AI.Confidence in autonomous and agenticsystemslooks at how autonomous AIsystems are changing our understandingof artificial intelligence. While manybasic ideas remain the same, autonomybrings new challenges that require a freshperspective.In this new era where autonomousAI systems interact and co-exist withhuman society, ensuring AI is reliable andmeets human expectations is of utmostimportance. This white paper expandsthese categories to address the specialneeds of autonomous and agenticAI systems.3 IConfidence in autonomous and agentic systems Table ofContents IConfidence in autonomous and agentic systemsMaking systemsautonomousLLMs within the agenticlandscapeUnderstanding agentpropertiesJourney tomulti-agency07051416 World models: Howagents understandtheir environmentConclusionBringing it alltogetherPurpose andalignment18272220 Journey to multi-agencydiagramAutonomous AI systems have existed long before today’s AI boom. For many years,systems with different levels of independence and integration have gradually developed.Understanding this history helps us appreciate the full gravity of today’s advancements.AutopilotChatbotCo-pilotMulti-agent systemAutomationIntegration5 IConfidence in autonomous and agentic systems Though the era of agentic AI began 50years ago, the development of simplechatbots showing limited independenceand integration have brought thetechnology into recent focus. These basicsystems, like early website help tools, hadlimited capabilities and limited accessto information which was not alreadyincluded in the models they used.As integration improved, co-pilot systemsemerged. These more advanced toolscould access and interpret data acrossmultiple systems, offering more helpfulassistance. However, for the most part,these systems only gave advice andcouldn’t act on their own.The next step forward was autopilotsystems. These tools had enoughindependence to take specific actions,but lacked the complete integrationcapabilities required to handle entireprocesses. These systems underlinedthe value of letting machines actindependently.Today, we’re seeing multi-agent AIsystems that combine high levels of bothindependence and integration. Theseadvanced systems can create greatvalue, but need careful oversight and riskmanagement. The move from single AIagents to systems with multiple AI agentsworking together marks a fundamentalchange in both the capability andcomplexity of this technology.This history shows an important fact:the shift from single agents to multi-agentsystems isn’t just about having moreagents; it’s about creating systems withentirely new properties and emergentbehavior when multiple agents worktogether in shared environments.6 IConfidence in autonomous and agentic systems Understanding agentpropertiesAgents are central to modern AIdiscussions, and terms like “agent” and“agentic” are used everywhere, often withvaryin