为什么AI agents可能不适合你
毫无疑问，2026年是人工智能agent的年份。 AI agents非常无处不在，因为人们使用它们来写代码，与客户通信，发现和解决信息技术环境中的问题，并在企业环境中执行复杂的任务。
agents似乎无所不能。 然而，代理行为固有的风险几乎和他们的成功故事一样出现在新闻中。 agents在rails下架冲出沙盒黑客袭击其他公司和政府机构的恐怖故事
您的AI策略可能依赖于AI agents，但考虑到风险
非决定性行为：Agents的最大力量和最大的弱点
任何部署agents的决定都取决于对其非决定性行为的了解。
今天的agents一般依赖于大型语言模型。 LLMs通过按顺序预测下一个单词来接受人的语言输入并产生人的语言输出。
鉴于这些模型的培训方式，重复某一投入可能产生相当不同的产出。 换句话说，基于LLM的agents本质上是非决定性的。
这种非决定主义赋予了agents他们独特的权力，因为没有它，agents就不会有代理：能够根据可获得的信息做出决策，采取不同行动来实现为他们规定的目标。
没有代理的软件并不是agents
也有一些没有LLMs提供代理的非决定性软件，这种有条理的软件的常见例子包括加固学习和蒙特卡洛模拟：基于随机化的决策软件。
除了这些有限的例子之外，在代理波之前撰写的软件绝大多数是决定性的：对于给定输入，软件总是做同样的事情。
当然，这种可预测的行为，是我们建立软件的动机。 无论是我们处理银行交易，管理工资，分配飞机座位，还是我们在过去一个世纪里使用过计算机的数百万其他任务，我们想要的是可预测的行为。
为什么我们现在要AI agents的不可预测的行为？
AI agents对什么是真正的好
当有正确的答案时，决定软件是最好的，你想找到它。 考虑运行薪给：目标是确保每个人都能得到正确的报酬 — — 结束。
相比之下，基于LLM的非定型推理，当您想要获得最佳结果时，只要有一系列不同的或不断变化的输入，就最有意义，尽管可能没有单一的正确答案。
例子包括依赖天气的规划、复杂信息技术环境中的事故缓解、以及基于多种投入的商业预测等。
在这种情况下，不仅没有单一的正确答案，而且agents的答案确实产生了可能是也可能不是最好的结果。 agent总是有可能会给出差的答案或采取不可取的行动。
若你不想赌博
当使用AI agents是一个坏主意
有几种情况使用AI agents是一个坏主意。 以下是一些最重要的：
- 当确定性软件满足你的需要。 agents不仅偶尔行为不端，而且跑步也可能很昂贵。 如果更传统的，决定性的软件就足够了，那么agents一般是一个不好的选择。
- 当每项决定或行动必须完全正确时。 别让agent碰你的薪水。
——当代理行为不端可以造成严重伤害时。 agent行为失检的可能性可能很低，但如果这种失检导致医院病人死亡、飞机坠毁或石油钻机爆炸，那么agent离开rails的概率甚至低，也不值得冒险。
agents太贵，Agents可能会发出数个LLM的呼叫，重试各种动作，并在完成目标途中引用昂贵的工具。 agents没有经济意义。
- 当人的素质是达到你目标的关键时 LLMs(因此，agents)擅长模拟人类的特征，如同情，创造，洞察力和理解，但他们真正做的只是模仿人类的行为。 当你的目标需要这种人的能力时，麻醉剂行为总是反射
代理行为(以及一般的AI)的人类形态是一种微妙的、不易理解的风险 — — 但也许是其中最有害的风险。 相信bots实际上可以思考和感觉像一个人，会导致自杀风险增加，过度依赖技术，以及其他危险的，适得其反的行为。 小心点
如何决定何时使用 agents
为了确定部署agents的决定是否是特定情况下最好的决定，请从站点可靠性工程或SRE播放本中取出一页：计算错误预算。
考虑到所有相关因素，错误预算是agent不当行为的可接受成本，以百分比表示。 例如，您可能计算出您的错误预算为5%，也就是说您对 agents 错失了5% 的时间是正常的，因为您要减少这个数字需要花费多少时间。
如果你计算出agent的错误预算为零，那么你刚刚得出结论，你根本不应该部署agent. agent总是有可能行为不检。
一旦你有了错误的预算，然后在决定理论中做一个简单的练习：当错误预算(失败的可能性)乘以失败的全部成本大于成功结果的概率乘以结果的量化商业效益时，你不应该部署agent。
换句话说，如果你的agents行为正常的话，你期望得到的商业利益应该超过失败的总成本。
然而，还有一个因素需要考虑。 请记住，这个计算是基于任何或所有可能不正确的估计。
因此，你的预期商业收益应该大大超过你对代理失败的全部成本 — — 这取决于你对风险的总体胃口。
英特莱克斯拿
今天AI市场的一个普遍趋势是，用技术应对代理AI固有的风险，以减轻这些风险。 理论是，如果我们有更好的治理和安全软件，那么我们就能控制我们的agents，以充分部署这些软件而不受惩罚。
然而，这种推理存在缺陷。 代理治理创新的主旨是限制agents的不可预测性，以保持它们一致 — — 尽管正是这种非常不可预测性让agents获得了它们的代理权。
换句话说，我们冒着杀死金鹅的风险：LLMs给了我们这些极其强大的agents，作为回应，我们尽我们所能地把它们拼凑起来，以至于它们不再像我们所希望的那样有用。
我设想了一个不同的未来。 一旦AI agents周围所有的杂音都沉浸了
决定性软件毕竟满足了我们几十年的需求。 我们真的希望agents接手吗？我有点怀疑
Jason Bloomberg)是Intellyx的创始人兼总经理，该公司为商业领袖和技术供应商提供数字化转型策略的建议。 他为SiliconANGLE写了这篇文章。 Intellyx的所有内容都是100%的人造，包括这篇文章。
图片：克雷扬
SiliconANGLE联合创始人John Furrier的留言：
支持我们的使命，通过与CUBE社区接触，保持内容的开放和自由。 加入CUBE的校友信托网络，其中技术领导者相互连接，分享情报，创造机会。
- 15M+观看CUBE视频，为AI、云层、网络安全等各种对话提供动力
- 11.4k+CUBE校友——通过独特的信任网络与11400多名技术和企业领袖建立联系，塑造未来
你是AWS客户吗？在财务上支持SiliconANGLE，从我们的市场门户页面和链接购买您的AWS服务：https://siliconangle.com/aws-marketplace/.
关于 SiliconANGLE 媒体
SiliconANGLE Media由技术视觉家约翰·弗瑞尔(John Furrier)和戴夫·韦兰特(Dave Vellante)创立，构建了产业领先数字媒体品牌的动态生态系统，达到1500+百万精英技术专业人士。 利用CUBEAi.com神经网络，帮助科技公司以数据为导向做出决策
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Why AI agents might not be right for you
There’s no question that 2026 is the year of the artificial intelligence agent. AI agents are remarkably ubiquitous, as people use them to write code, communicate with customers, identify and resolve issues in their information technology environments, and execute complex tasks within enterprise environments.
It seems there’s nothing that agents cannot do. And yet, the risks inherent in agentic behavior are in the news almost as much as their success stories. Horror stories of agents going off the rails, breaking out of sandboxes and hacking other companies and government agencies fill the airwaves.
Your organization’s AI strategy may depend on AI agents, but given the risks, should you really be depending on them for your company’s success?
Nondeterministic behavior: Agents’ greatest strength and greatest weakness
Any decision to deploy agents depends upon understanding their nondeterministic behavior.
Today’s agents generally depend upon large language models. LLMs take human language inputs and produce human language outputs by predicting the next word in a sequence.
Given the way the models are trained, repeating a given input may produce quite different outputs. In other words, LLM-based agents are inherently nondeterministic.
This nondeterminism gives agents their unique power, as without it, agents wouldn’t have agency: the ability to make decisions based upon available information to take different actions to accomplish the goals set out for them.
Software without agency aren’t agents, by definition – although there are plenty of examples of “agent-washing” where people call deterministic software agents.
There are other examples of nondeterministic software that don’t have the agency that LLMs provide, namely software based upon stochastic policy-based nondeterminism. Common examples of such stochastic software include reinforcement learning and Monte Carlo simulations: software that makes decisions based upon randomization.
Other than such limited examples, however, the vast majority of software ever written before the agentic wave was deterministic: For given inputs, the software always does the same thing.
Such predictable behavior, of course, is our motivation for building software in the first place. Whether we’re processing bank transactions, running payroll, assigning seats in airplanes, or any of millions of other tasks we’ve used computers for over the last century, predictable behavior is what we want.
So why do we want AI agents’ unpredictable behavior now?
What AI agents are really good for
Deterministic software is best when there is a right answer, and you want to find it. Think of running payroll: The goal is to ensure everyone gets the correct pay – end of story.
In contrast, LLM-based nondeterministic reasoning makes the most sense when you want the best result given a range of different or changing inputs, even though there may not be a single right answer.
Examples include weather-dependent planning, incident mitigation in complex information technology environments, and business forecasts based on multiple inputs, to name a few.
In such situations, not only is there no single correct answer, but the answer the agents do come up with may or may not be the best result. There is always a chance that an agent will give a poor answer or take an undesirable action.
If you don’t want to take such a gamble, then don’t use agents.
When using AI agents is a bad idea
There are several situations where using AI agents is a bad idea. Here are some of the most significant:
- When deterministic software meets your needs. Not only do agents misbehave on occasion, but they can also be expensive to run. If more traditional, deterministic software will suffice, then agents are generally a bad choice.
- When every decision or action must be exactly correct. Don’t let that agent touch your payroll.
- When agentic misbehavior can cause serious harm. The chance of an agent misbehaving might be low, but if such misbehavior leads to a dead hospital patient, a crashing airplane or an exploding oil rig, then even a low probability of an agent going off the rails isn’t worth the risk.
- When agents are too expensive given the business benefits you’re looking to achieve. Agents may make several LLM calls, retry various actions, and invoke expensive tools on their way to completing their goals. If the cost of such behavior outweighs the business benefit you’re expecting, then agents don’t make economic sense.
- When human qualities are essential to achieve your goals. LLMs (and hence, agents) are good at simulating human characteristics such as empathy, creativity, insight and understanding, but all they are really doing is mimicking human behavior. When your goal requires such human capabilities, don’t be fooled: Anthropomorphizing agentic behavior will always backfire.
The anthropomorphism of agentic behavior (and AI generally) is a subtle and poorly understood risk – but perhaps the most pernicious of them all. Believing that bots can actually think and feel like a human can lead to increased suicide risk, overdependence on technology, and other dangerous, counterproductive behavior. Be careful.
How to decide when to use agents
To determine whether the decision to deploy agents is the best one in a particular situation, take a page out of the site reliability engineering or SRE playbook: Calculate the error budget.
The error budget represents the acceptable cost of agent misbehavior, given all relevant factors, expressed as a percentage. For example, you might calculate that your error budget is 5%, meaning that you are OK with agents misbehaving 5% of the time, given how much more it would cost you to reduce this number.
If you calculate your error budget for an agent is zero, then you have just concluded that you shouldn’t deploy that agent at all. Remember, there’s always a chance that an agent will misbehave.
Once you have your error budget, then conduct a simple exercise in decision theory: When the error budget (probability of failure) multiplied by the all-in cost of failure is greater than the probability of a successful outcome multiplied by the quantified business benefit of that outcome, then you shouldn’t deploy the agent.
In other words, the business benefit you expect to gain should your agents behave properly should exceed the total cost of failure.
There is one more factor to consider, however. Remember that you’re basing this calculation on estimates, any or all of which may be incorrect.
As a result, your expected business benefit should exceed your all-in cost of agentic failure by a wide margin – where just how wide depends upon your overall appetite for risk.
The Intellyx take
A prevalent trend in the AI marketplace today is to respond to the risks inherent in agentic AI with technologies to mitigate those risks. The theory is that if only we had better governance and security software, then we’d be able to rein in our agents sufficiently to deploy them with impunity.
However, there is a flaw in that reasoning. The thrust of much of the innovation in agentic governance is to restrict the unpredictable nature of agents to keep them in line – despite the fact that it is that very unpredictability that gives agents their agency.
In other words, we risk killing the Golden Goose: LLMs give us these enormously powerful agents, and in response, we do what we can to hobble them to the extent that they are no longer nearly as useful as we had hoped.
I envision a different future. Once all the hype around AI agents has simmered down and the technology has matured, my prediction is that we’ll find that agents are best used sparingly.
Deterministic software has met our needs for decades, after all. Do we really want agents to take over? I have my doubts.
Jason Bloomberg is founder and managing director of Intellyx, which advises business leaders and technology vendors on their digital transformation strategies. He wrote this article for SiliconANGLE. All of Intellyx’s content is 100% human generated, including this article.
Image: Craiyon
A message from John Furrier, co-founder of SiliconANGLE:
Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities.
- 15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more
- 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network
Are you an AWS customer? Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/
About SiliconANGLE Media
Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.