Agentic AI: The Next Frontier is Cross-Functional Collaboration Between Domains

Varun Uppal, the Founder and Chief Product Officer at Shinobi Security, says, despite rapid progress, foundational challenges like hallucinations and brittle reasoning still exist

How do you define Agentic AI, and how does it fundamentally differ from traditional chatbots or rule-based automation?
Agentic AI is designed to autonomously complete entire tasks, not just assist with isolated steps. Traditional chatbots might provide information, for example, answering: “What are the top 5 cheapest flights to London?” But Agentic AI goes further. It can take that intent and execute the full workflow “Book me the cheapest direct flight to London on July 20th, 2025.”

The key difference is that Agentic AI doesn’t just answer questions, it takes action. It uses AI models connected to tools and other specialized agents, working together to handle complex tasks. This allows it to plan, execute, and adjust its actions to achieve a goal, much like a skilled human assistant.

Beyond task automation, what complex, multi-step problems are agentic AIs uniquely positioned to solve that current AI models cannot?
Research is a great example. Whether it’s evaluating the pros and cons of a specific car model or conducting sophisticated scientific or financial analysis, research is inherently multi-step and iterative. Each new piece of information leads to further questions, often resulting in dozens of articles, papers, or datasets that need to be reviewed, synthesized, and acted upon.

Agentic AI shines in these situations because it can handle the entire research process on its own. Starting with a user’s goal, it can search multiple sources, gather and analyze information, write code to process data, and adjust its approach as it learns more. Instead of just providing information, it actively works through the problem to deliver useful insights.

How much autonomy should Agentic AI have in decision-making, and where should humans remain in the loop?
The degree of autonomy should be proportional to the risk. For low-stakes or fully reversible tasks, agentic systems can safely operate end-to-end. But in any scenario where the consequences of an action carry significant risk – financial, legal, operational, or reputational – humans should remain in the loop until there’s sufficient confidence in the system’s reliability.

Take AI penetration testing as an example: an agent might find vulnerabilities and suggest ways to exploit them, but any risky action, like deleting or changing data, should always need human approval. The same applies when agents produce insights for business decisions; people still need to review both the results and how the agent reached them.

Despite rapid progress, foundational challenges like hallucinations and brittle reasoning still exist. Until these are fully addressed, human oversight isn’t just a safety net, it’s a necessary part of responsible deployment.

What infrastructure challenges (such as power, cooling, latency) arise when deploying Agentic AI at scale?
The infrastructure challenges depend on how complex and long-running your agents’ tasks are. If you’re running your own models, the first bottleneck is compute, specifically access to GPUs or specialized chips that can handle heavy, sustained workloads. Unlike chatbots that only use compute briefly when responding to a question, agentic systems often run complex tasks that can last minutes, hours, or even days. This constant activity drives up power and cooling costs and requires infrastructure that can handle high loads for long periods.

Before scaling agentic systems, it’s essential to run controlled experiments to establish baseline resource requirements, cost models, and failure points.

What’s the next frontier for Agentic AI—will we see AI “agents” collaborating like human teams?
We’re already seeing early versions of this. Many agentic platforms are structured as swarms of specialized agents collaborating with one another, often inspired by how human teams operate. For instance, an agentic accounting system might mirror a corporate finance department, with dedicated agents for accounts payable, accounts receivable, compliance, and reporting, each interacting to manage complex workflows.

The next frontier is cross-functional collaboration between domains. Imagine your accounting agents working alongside HR agents, legal agents, or supply chain agents, negotiating tradeoffs, resolving conflicts, and coordinating actions just as human teams do. As models become more capable of reasoning, negotiating priorities, and maintaining shared context, we’ll move closer to true digital organizations.

Will open-source models keep pace with proprietary Agentic AI systems, or will there be a widening gap?
This is one of the most dynamic and difficult areas to predict. Fundamentally, progress hinges on discovering new methods to advance model capabilities. Proprietary players enjoy a structural advantage here: they can combine internal research breakthroughs with innovations from the open-source community, benefiting from both closed and open ecosystems. This allows them to compound their lead.

Open-source models often close the gap when progress relies on scaling existing techniques or refining well-understood methods. But when breakthroughs require novel architectures, training paradigms, or proprietary data assets, the gap tends to widen. At the moment, we’re likely to see proprietary models maintain their lead, at least until the next major wave of open research innovation levels the field again.