The Rise of AI Agents: Transforming Business and Technology Landscapes

From co-pilots to AI agents to custom tool handling - what should marketers do next?

AI has been the topic of most conferences happening this fall, and listening to some speakers, there are some exciting possibilities we need to take advantage of. But we also need to be careful. Every software vendor rushed to implement a conversational chatbot or co-pilot as soon as OpenAI announced ChatGPT. Most of them revolved around helping perform content improvements. Vendors are still adding co-pilots to their applications to varying degrees of usefulness. For instance, HubSpot announced an updated co-pilot at Inbound that works across the entire platform. It's an improvement on what they had before, but it still needs some work. Christopher Penn, Chief Data Scientist and co-founder of Trust Insights, an AI consulting firm, said at the recent Martech Conference that co-pilots will no longer be a differentiating feature in software. It doesn't matter how good co-pilots are or aren't, AI agents have taken over much of the conversation and development effort. At the Inbound Conference, HubSpot Founder and CTO Dharmesh Shah said 2024 is the year of AI agents. HubSpot defines an AI agent as "software that uses AI and tools to accomplish a goal requiring multiple steps." These goals can be simple or complex. In either case, they share three things: HubSpot announced a set of AI agents for its customer platform, including a social media agent that can create a full social media calendar and a content agent that can take a podcast and build assets like a summary blog, clips, and more. Salesforce also announced plans for Agentforce at Dreamforce, including allowing Agentforce to connect to other AI platforms. Plus, other companies are either announcing new agents or talking about them (and some already have them). In his talk, Shah said AI models are improving faster and innovating on multiple dimensions. That means agents can take on more sophisticated goals. He shared some examples of agents on Agent.ai, a professional network for AI agents that he created. One AI agent can do market research on a company, pulling in all the information they can find across a range of categories. Others can analyze earning calls, optimize content for SEO, and generate email subject lines. Shah said we need to think of agents as digital teammates and stop thinking AI will replace us. Instead, AI raises your value, not reduces it, he said. "Humans are more than the sum of the tasks that we do." We need to allow AI to automate the mundane and amplify the magic that makes us us. Shah offered three predictions about AI agents: While agents are the focus right now, they aren't the only way martech companies should leverage generative AI. During his session at the Martech Conference, Penn discussed how large language models have worked around their text-only limitations. LLMs perform text-based language tasks. They cannot work with images, audio, or video. But you see today that ChatGPT is creating images, and you can upload images, audio, and video and ask the LLM to create summaries, analyze images, and perform other tasks. Other LLMs are doing similar things. So, how are they doing it? They are using tools. Penn explained that LLMs can't do math, but they can recognize a math problem and write code that does the math. The LLM assembles the code, runs it, and returns the answer. He said writing code is a language problem, and it's implicit in an LLM's architecture. All the LLMs leverage code execution, which allows them to work with content and data that is not text-based. Penn shared another example: Meta's Llama 3.1 has built-in tool handling. Not only can it write its code, but it also knows to look for an API to allow it to search the web, including tying into Wolfram Alpha (a more sophisticated math engine). Custom tool handling is an opportunity for software vendors, Penn said. He said if a martech vendor can expose its services to the models, then the models can intuit when to use that tool. For example, a model can build an API to connect to Demandbase to pull in intent data. Or it could connect to Zoominfo to pull in a contact's information. This goes beyond creating GPTs for ChatGPT (which is a type of tooling but only works for one vendor). Penn said that other models, like Gemini, Claude, and Llama, also have function calling, which means these models can call multiple services to achieve a goal requested by a person. Penn believes martech vendors should expose their applications to the models and provide coders and models with the necessary sample code to show them how to do it. And they should be marketing this feature because it doesn't require extra work on the customer's part other than needing a subscription to the vendor's service. Why is this important? It enables software vendors to generate new sources of revenue. For example, a company might not want to work with a new application but wants access to some of its capabilities or data from within a model they are using in another application. This is similar to the idea of agents connecting to other agents. Martech vendors that do this will jump ahead of the rest. I've always been somewhat bored with the discussions around generative AI and content development because I could see the possibilities around using generative AI to automate tasks. AI agents and AI tooling speak to what I wanted to see us talking about because the value is so clear. Helping marketing and sales teams do their work more efficiently empowers them to focus on important aspects of their work that AI can't do. But it's also starting to get a little crazy with the feeling that every martech product or platform will create one or more AI agents for their platform. Like the GPTs that only work with one vendor, AI agents that only work with a single platform may have value for that platform, but it's not enough. We need to learn from the challenges companies face with a disconnected martech stack. Platforms like HubSpot that have marketplaces must build agents that can work with applications in those marketplaces. Shah talked about developing an Agent User Interface (AUI) allowing agents to talk to each other, but it's experimental right now. All martech vendors need to build agents that can talk to applications they integrate with (and vice versa), whether that's other agents or through custom tooling for LLMs. If martech (and sales tech) vendors aren't thinking holistically, we're in for a new set of challenges.

Meta, Outshift, Intuit and Asana dig into the agentic AI future

Kumar Sricharan VP Technology & Chief Architect for AI, Intuit; Shubha Pant VP, AI/ML, Outshift by Cisco While business and technology leaders are working to reap value from AI today, the future is barreling toward us, and it's vital to build a foundation for what's quickly coming. Tech leaders are betting that the future of AI is agentic. In other words, organizations will adopt intelligent systems that not only perform tasks autonomously, but also make decisions with near-human-like precision, from writing code to handling ecommerce operations, functioning as automated sales agents and more. Agentic AI was the focus of the VentureBeat AI Impact Tour: "Agentic AI -- the next giant leap forward in the AI revolution," presented by Outshift by Cisco. Speakers from Outshift, Meta, Asana and Intuit joined VB CEO Matt Marshall to explore how orgs can plan for an agentic AI future and other onrushing advances in AI. Building out an agentic future "If you think about that future where these agentic systems are going to work together to solve bigger problems, we need distributed agentic systems computing and we need an open, interoperable internet of agents," Vijoy Pandey, GM and SVP, Outshift by Cisco told Marshall. "Innovation slows down when you're in a walled garden. Whether you're an infrastructure vendor, operator, an app developer and most importantly a consumer or customer, an open system provides value for each individual link in the chain." Agentic systems that learn how talk to each other and interconnect have the power to change the way humans work, starting with software and IT and then moving toward knowledge work, services and even physical work as robotics evolve. They'll also need to be integrated into existing software systems and physical environments, as well as instantiated on those existing software systems, whether it's cloud, on-prem or embedded in a robotic solution. Tying it all together requires abstraction layers. That will look like open models, open tooling, an orchestration and discovery layer and then a communication layer that is secure, stable and open. Then there's handling probabilistic outcomes, communication through NLP, the exchange of state information and more. "These could be pretty massive problems to solve," Pandey said. "We're looking for these problems and what they look like and how to solve them. That's where the future is." The time to start on AI agents is now Today's big question is whether the technology is mature enough to realize its full potential -- and it isn't quite yet. However, that can't be a barrier, Mano Paluri, VP of generative AI engineering at Meta, said, during a one-on-one conversation with Marshall. "You can't wait," he added. "In that sense, I would say that it's ready. Agents clearly feel like the next step in the evolution of these models. The way we have been thinking about it is moving away from a model to a system that has multiple components that are customizable." In the hunt for autonomous systems that can foresee, learn, reason, act and iterate to solve a complex problem, we've already come far in perception -- foundational models are able to learn from text and images. We're still in the early stages of reasoning through complex problems, but today models can learn at far larger scales than ever before over the last decade. And these models are beginning to plan, from both an inner and an outer loop perspective. Today the outer loop is the human training the model. Next will be the agent handling parameters itself. The Meta AI agent Today's Meta AI agent is the first step in the evolution of LLMs as Meta moves away from a model to a system that has multiple, customizable components. The goal is to fine-tune the model for every use case, extend the context window, adapt to a new language and so on, for all four billion customers. "We also believe in a family of agents," he said. "This incarnation of Meta AI is a user assistant, but we also think everyone should be able to customize the agent in the way that they want. This is the family of agents where businesses can create a billing agent. Creators can have their own agent to reach a larger scale. Advertisers can have creation capabilities that are unprecedented." Agentic AI use cases: challenges and opportunities To close out the night, Paige Costello, head of AI at Asana, Shubha Pant, VP, AI/ML at Outshift by Cisco, Kumar Sricharan, VP of technology and chief architect for AI at Intuit, joined Marshall for a conversation about the use cases agentic AI will open up, and the challenges and opportunities that will come hand in hand. Real-world case studies Handling requests within a workflow can be a huge time suck, but that's where agentic AI comes in. Asana has embedded agentic AI for both chat and workflow use cases. In the case of workflows, it can handle a request at the outset, determining where to prioritize it, whether there's enough information to get started and who should be included. It's a great place for a company to start adding agentic workflows, Costello added. "There are so many opportunities where AI can be a partner in doing this work," Costello said. "The agentic piece is, how much decision-making or autonomy does it have to do these things in the context of those workflows? We've seen great success with security companies, with marketing agencies. There are many other use cases where we're starting to see things like creative requests, working through revisions and feedback and approval loops." At Intuit, they're automating across their entire suite of financial products, and offering insights and financial guidance. Across that ecosystem of tasks, they're experimenting with AI, especially in areas where building hand-engineered solutions would be time-consuming, or even just fizzle. For instance, small businesses have a broad array of characteristics and needs. During onboarding, the customer is required to detail all that information so that Intuit can classify them. "Now, with the rise of agentic AI, we're finding that we can use these systems, these different agents that can work in unison to allow the customer to give us access to the different sources of their information," Sricharan said. "Then these agents essentially autonomously operate on top of that information and help the customer onboard with minimal effort on their part." Internally, agentic AI is helping the company navigate tax code changes that impact products. Instead of a team of developers researching and implementing new elements, they can use agentic AI there to span a variety of functions, all the way from detecting where the changes are, to making associations with our code and determining what changes need to be then made, acting as a copilot for developers. Outshift has an incubation team focused on building a multi-agent predictive diagnostic and remediation tool for enterprises across their tech stacks. The goal is to predict IT issues before they arise, identify root causes when issues occur and offer mitigation strategies until full resolution. There are other agentic AI projects in progress including architecture development, open standards for orchestration of agents, composition of multi agent systems, and an open agent protocol for inter-agent communication. "The main challenges for agentic AI right now are threefold," Pandey said. "First, how AI agents discover each other and understand each others' capabilities. Second, how they collaborate to solve problems and handle uncertain outcomes. And third, how they communicate using imprecise natural language instead of fixed structures like traditional APIs." "We need to figure out how to create standards and open-source guidelines for these AI systems that deal with probabilities," he added. "It's time for the tech community to join forces and build these solutions together."

The journey to fully autonomous AI agents and the venture capitalists funding them

The term "agentic AI," or "artificial intelligence agents," is rapidly becoming commonplace, so much so that those invested in the technology see a need to draw distinctions. In a series of blog posts published last week, partners at venture capital firm Menlo Ventures, (which has bankrolled startups in artificial intelligence such as Anthropic), define "the next wave of agents" and how they surpass the agents introduced so far. Tomorrow's agents, they write, have four distinct capabilities. Also: Networks of collaborative AI agents will transform how we work, says this expert "Fully autonomous agents are defined by four elements that, in combination, ladder up to full agentic capability: reasoning, external memory, execution, and planning," write the authors. "To be clear, the fully autonomous agents of tomorrow might possess all four building blocks, but today's LLM apps and agents do not," they declare. The authors, Tim Tully, Joff Redfern, Deedy Das, and Derek Xiao, explore in their first blog post what it means for something to be "agentic." The software, they write, must ultimately gain greater and greater autonomy in selecting between possible steps to take to solve a problem. Also: Bank of America survey predicts massive AI lift to corporate profits "Agents emerge when you place the LLM in the control flow of your application and let it dynamically decide which actions to take, which tools to use, and how to interpret and respond to inputs," the authors write. A conventional large language model can have access to "tools," such as external programs that let the LLM perform a task. Anthropic has already done this with its Tool Use feature, and OpenAI has something similar. However, the authors explain that invoking a tool merely gives an LLM means to solve a problem, not the control to decide the way a problem should be solved. Also: 98% of small firms are using AI tools to 'punch above their weight' As the authors write, "Tool use is powerful, but by itself, [it] cannot be considered 'agentic.' The logical control flows remain pre-defined by the application." Rather, the agent must have a broad ability to choose which tool will be used, a decision logic. A few versions of software come closer to being true agents, the authors explain. One is a "decisioning agent," which uses the large language model to pick from among a suite of rules that in turn decide which tool should be used. They cite healthcare software startup Anterior as an example of such a decisioning system. Next, a higher-order agent, called an "agent on rails," is "given higher-order goals to achieve (e.g., 'reconcile this invoice with the general ledger,'" they write. The program is granted more latitude to match the high-level request and which sets of rules to follow. Also: There are many reasons why companies struggle to exploit generative AI, says Deloitte survey Multiple startups are pursuing this "agent on rails" approach, the authors note, including customer service firm Sierra and software development firm All Hands AI. The third, highest level of agentic AI, the holy grail, as they put it, has "dynamic reasoning" and a "custom code generation" that allows the large language model to "subsume" the rulebook of the company. This kind of approach, known as a "general AI agent," is still in the research phase, the authors note. Examples include Devin, the "first AI software engineer," created by startup Cognition. In the second blog post, "Beyond Bots: How AI Agents Are Driving the Next Wave of Enterprise Automation," the authors reflect on how agentic AI will be applied in enterprises. The immediate impact, they write, is to move beyond "robotic process automation," or RPA, tools that replace some basic human tasks with software, sold by firms such as UiPath and Zapier. Also: 73% of AI pros are looking to change jobs over the next year The decision agents and agents on rails explored in the first post find practical applications in business tasks, such as reconciling supplier invoices to a general ledger: Let's say a company needs to reconcile an invoice from an international supplier against its ledger. This process involves multiple considerations, including invoice currency, ledger currency, transaction date, exchange rate fluctuations, cross-border fees, and bank fees, all of which must be retrieved and calculated together to reconcile payments. Agents are capable of this type of intelligence, whereas an RPA agent might just escalate the case to a human. The main thrust of the blog post is that numerous startups are already selling things that approach such higher agentic functions. They "aren't just science fiction, either," they write. "Although the category is still emerging, enterprises from startups to Fortune 500 companies are already buying and leveraging these systems at scale." Also: How to level up your job in the emerging AI economy The authors offer a handy chart of the numerous offerings, organized by the degree of autonomy of the agent programs along one axis, and the degree of vertical or horizontal-market focus: Not covered in the two blog posts are two key limitations that have cropped up in existing generative AI (gen AI) systems and threaten to stymie the progress of agents. First, there is no substantial discussion by the authors on how to deal with hallucinations, confidently asserted false output. Whatever the reasoning process used by gen AI, and however formidable the tools, there is no reason to suppose that AI agents won't still generate erroneous outputs like conventional chatbots. Also: Prepare for AI-powered 'agent ecosystems' that will dominate tomorrow's services At least, the question of whether or not decision agents and agents on rails diminish hallucinations is an open research question. Second, while agentic AI can conceivably automate a number of corporate processes, there is to date very little data on the effect of that automation and whether it is truly an improvement. That is partly connected to the first point about hallucinations, but not entirely. An agent that is not wrong in its reasoning or actions can still lead to outcomes that are suboptimal versus what a person would do. A prominent example is discussed in the book, "AI Snake Oil" by Princeton computer science scholars Arvind Narayan and Sayash Kapoor, published this month by Princeton University Press. An AI model tracked the history of patients with asthma who presented with symptoms of pneumonia when entering the hospital. The AI model found they were among the patients with the lowest risk in the hospital population. Using that "reasoning," such patients could be discharged. Also: Asking medical questions through MyChart? Your doctor may let AI respond Yet, the model missed the causal connection: patients with asthma and symptoms of pneumonia were least risky because they received emergency care. Simply discharging them would have bypassed such care and the results could have been "catastrophic," Narayan and Kapoor declare. It's that kind of correlation instead of causality that can lead to vastly sub-optimal results in real-world situations with complex causal situations. Also left out of the authors' scope of discussion are agents that collaborate. As Hubspot CTO Dharmesh Shah told ZDNET recently, the future work of agentic AI will not be done by a single agent but likely by networks of AI agents collaborating with one another. Also: AI is relieving therapists from burnout. Here's how it's changing mental health Given those omissions, it's pretty clear that despite the sweep of the venture capitalists' research, they have only scratched the surface of what will be achieved in a world of increasingly powerful AI agents.

Networks of collaborative AI agents will transform how we work, says this expert

It's commonplace for artificial intelligence (AI) experts in research and industry to talk about AI agents and agentic AI as the focus of innovation. OpenAI's Sam Altman said as much almost a year ago at the company's unveiling of an online store for custom GPTs and has elaborated at length about how agents will take on much more complex tasks than today's copilots. The next step beyond today's agentic AI may be networks of agents that collaborate to achieve more complex tasks, mostly without human supervision, according to Dharmesh Shah, co-founder and chief technology officer of software maker Hubspot. Also: AI agents are the 'next frontier' and will change our working lives forever "I think over the fullness of time, we're gonna see these agents collaborate with each other, right?" said Shah in an interview with ZDNET via Zoom this week. "Agents are effectively a progression up from copilots," said Shah. "And I think both will have their place in the business landscape. What makes agents interesting is that they can take on, kind-of, higher-order goals that usually involve multiple steps." Also: Bank of America survey predicts massive AI lift to corporate profits Hubspot is competing with firms, such as Salesforce, to roll out various agents to assist with customer relationship management (CRM) tasks, including sales, marketing, customer relationship management, and more. To connect those agents, Hubspot is promoting the idea of a network that acts as a marketplace for agents. At its annual user conference a week ago, Inbound, alongside a suite of AI offerings call Breeze, Hubspot unveiled a network for agents called agent.ai. The offering has over 47,000 users, Hubspot announced, and more than 1,700 builders signed up to create agents of their own. "This is a professional network for agents, which I am personally involved in," Shah told ZDNET. "Think of it as the number one professional network for agents, unlike LinkedIn, which is for humans." Also: Prepare for AI-powered 'agent ecosystems' that will dominate tomorrow's services Through the network, you can imagine teams of agents, consisting of "mini agents", and a supervising agent, explained Shah. "Over time, as these agents develop, they're going to be able to, kind-of, use each other. So, one agent says, 'I'm an agent that helps you do research on a company,' right?" That agent would look through public company transcripts, such as earnings calls. "Then, there's an agent that will go look at the [corporate] website, and see how web traffic and all those kinds of things are doing, it'll pull all that data together," said Shah. Also: 3 ways to build strong data foundations for AI implementation "So, these global kinds of mini agents are used by a kind of higher-level agent, so, you're sort-of composing these agents like Legos, then building higher-order structures." The agents become "digital teammates", according to Shah. The agents.ai network is a marketplace to find which agents can do which tasks, "see what their experience is, whether they'd be a good fit or not" based on the feedback of users of agents. Past efforts to build collaborating "objects", such as the CORBA standard in the 1990s, were stymied by interoperability, noted Shah. This time around, the natural language ability of generative AI becomes the connective tissue for programming and assembling agents. Also: Salesforce unveils AI agents for sales teams - here's how they help "We can interact with AI through natural language, right? Well that also carries over to agents," said Shah . "The API, so to speak, of an agent is actually natural language. You don't have to learn this other language to be able to make use of these agents, whether you're doing it as a human or you're doing it as another agent. That unlocks a new level of interoperability that I think has historically been hard to accomplish." Shah said the greater significance of agentic AI, and networks of collaborating agents, is a reinvention of CRM software. "We saw lots of innovation back in the cloud days," said Shah, referring to the 2006 to 2007 time frame when cloud computing first burst on the scene, "which was the last big kind of 'transformative impacts every industry kind of thing' -- you know, cloud CRMs emerged, and have taken over now pretty much every kinda major CRM." Also: A third of all generative AI projects will be abandoned, says Gartner Today's "paradigm shift" in CRM will be AI-led. "We're going to have now a new kind of paradigm, which is going to be an AI-based smart CRM," said Shah. That shift means the competitive battle in CRM, between Hubspot and Salesforce and others, said Shah, will be about which "platforms" provide the best use of agents for both users and developers. "One of the reasons I'm personally very excited about this, kind-of, AI transition, is that now we're going to see a whole new generation of developers that are going to be seeking a platform on which to build their ideas, saying, 'Hey, I want to build something for marketing or sales,'" he said. "Which CRM platform will they choose? And, now, I think we have an opportunity to build this kind of new mindshare within the developer ecosystem to say, 'Hey, now it's not about Web-based applications; agents are the new apps,' right? That's the new thing that people will be building. So, that's exciting."