Presented by Outshift by Cisco
Gartner has predicted that by 2028, one third of human interactions with generative AI will evolve from users prompting large language models (LLMs) to users interfacing directly with autonomous, intent-driven agents – a major jump ahead of the reactive AI assistants many users are now familiar with.
“Agents are the next evolutionary step in generative AI,” Vijoy Pandey, SVP/GM of Outshift, Cisco’s incubation arm, told VentureBeat. “For the C-suite, the message here is that you need to be prepared. It’s just three years away. There is a long journey ahead, so you need to start today with assistants, with low hanging use cases that don’t have a blast radius which is massive, and then slowly and steadily move towards use cases that are more critical.”
Think of AI agents as tireless, specialized employees in an organization that are very specifically tailored to a task.
You can think of AI agents as tireless, specialized employees in an organization that are very specifically tailored to a task, and collaborate to solve business problems for you, Pandey adds. Adoption is currently picking up steam, and showing great results, Tim Tully, partner at Menlo Ventures, told VentureBeat.
“I’m seeing a remarkable stream of customer success companies replacing and augmenting customer success teams with agents and helping them scale out,” Tully said. “It’s happening in marketing automation. It’s happening in code generation. I think you’re going to see agents spread across into other forms of software engineering as well. They’re incredibly pervasive as it stands today, but I think in the future agents are going to be used even more broadly across the enterprise.”
The Big Three (see Google Cloud, Microsoft’s Copilot stack and Q from AWS) are all getting into the game and building generative AI agents – and that is usually a pretty solid cue that an important new technology is on the playing field.
What sets agents apart from assistants
What’s the big difference – and what sets AI agents apart from the previous generation of AI-powered assistants that many users have gotten comfortable with?
AI assistants are essentially tools that react to user requests using LLMs and natural language processing (NLP). Users can ask questions or make requests, and the assistant hunts down answers and generates contextual content in a conversational interface.
Agents are proactive and autonomous, making decisions and taking actions without the need for human intervention.
AI agents, in contrast, are interactive but more importantly, proactive and autonomous, making decisions and taking actions without the need for human intervention. They’re always online, listening, reacting and analyzing domain-specific data in real time, making informed decisions and acting on them. They are designed to handle complex end-to-end workflows without supervision, but are usually created to tackle specific tasks and work toward specific goals.
And unlike their predecessors, agents turn out high-quality content, which reduces review cycle times by 20 to 60% — and can show their work, since the chain of tasks and data sources can be easily accessed and reviewed.
“Think of them as tireless, specialized employees in an organization that are very specific to certain tasks and that collaborate together to solve a bigger business problem for you,” Pandey said. “The big difference would be that you would not ask your CFO about a marketing campaign. You would not ask the same agent about 20 different things. You would want these things to be really specialized so that they can be accurate, and you want them to come together to solve a problem for you.”
In financial services an agent could detect and prevent fraud as it’s happening. It could handle dynamic financial planning with real-time budgeting, forecasting and scenario analysis. In HR, an agent could analyze candidate data to identify top talent, predict employee turnover and provide personalized career development recommendations. In marketing, AI agents could continuously analyze campaign performance, making real-time adjustments for maximum ROI, or continuously monitor competitors’ activities, identifying opportunities and threats.
When agents are integrated into a multi-agent framework, the resulting systems can collaborate across skill and knowledge areas using a variety of protocols and communication channels, understand data pulled from multiple sources, make decisions and handle more complex workflows from start to finish, without the need for human intervention or control.
However, an agent-specific orchestration layer that fully supports that evolution hasn’t been developed yet, Tully says, and that’s a major opportunity for startups.
“There needs to be some kind of Kubernetes-like infrastructure that allows agent tech workloads to run on, somewhere between lambda functions or ephemeral functions and Kubernetes.“
“There needs to be some kind of Kubernetes-like infrastructure that allows agent tech workloads to run on, somewhere between lambda functions or ephemeral functions and Kubernetes — a Goldilocks solution of not too hot, not too cold,” he explains. “It’s thin agents that are great at specific tasks, that are networked together, talking to each other with some protocol that’s yet to be devised, a multicast broadcast across the substrate so they’re able to communicate seamlessly with each other.”
Making the leap from assistants to agents
The Cisco AI Readiness Index found that 97% of organizations want to leverage generative AI, but only 14% already do it – that’s a huge gap that needs to be bridged, before an organization can even consider leveraging AI agents. There are a number of barriers just at the generative AI level. Common challenges include not knowing where to start, delivering ROI, and managing the unique trust, safety and security challenges that generative AI brings. There are also larger, industry-spanning challenges, Pandey said, from LLM hallucinations to successive loops in agents, where the decision-making workflow gets stuck because outputs are rejected.
“If you throw an ambiguous problem or a high-level problem statement at an agent, you need internal reasoning and planning to build a set of instructions that they can go and tackle,” he explained. “While they can build relationships, it’ll take us a while to get to good enough reasoning to prevent loops.”
“Start with a mundane business case instead of a moonshot.”
Organizations need to start somewhere, however. That includes empowering the citizen developers, especially those in the business functions that can most benefit from generative AI solutions – the ones who deeply understand the business processes and procedures in their areas and how to improve them. This is especially key given how scarce on-the-ground generative AI developers still are.
A foundational step before organizations can start their AI journey with AI assistants, and in the future agents, is data cleansing. Organizations should make sure to fix data hygiene issues and get identities and access control in shape.
“Start with a mundane business case instead of a moonshot,” Pandey said. “This allows you to go through the entire process of building out that pipeline and educating citizen developers. Then you can build the foundation up. AI is a journey, and the good news is that if you solve for the problems of generative AI with assistants in your organization, you will then end up solving for the agent world as well.”
Fortunately, the problems of AI occur the same across use cases and are well on their way toward being solved. As more industries make the leap from assistants to agents and as LLMs get better, every organization will reap the benefits and be well-positioned to take full advantage of the benefits of agentic generative AI.
Watch the whole conversation with Outshift’s SVP/GM Vijoy Pandey, Tim Tully, partner at Menlo Ventures, and VB editor-in-chief Matt Marshall here.
Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact
