Christian Reilly, the Field CTO for EMEA at Cloudflare, discusses the transformative potential of Agentic AI. Reilly explains how this advanced form of artificial intelligence moves beyond traditional chatbots and rule-based automation, enabling true autonomy and proactive problem-solving
How do you define Agentic AI, and how does it fundamentally differ from traditional chatbots or rule-based automation?
At Cloudflare, we see Agentic AI as a fundamental leap beyond traditional AI. Think of a traditional chatbot: it’s largely reactive, confined to pre-programmed scripts and responding within a defined conversational context. It’s excellent for answering FAQs or guiding users through set processes. Similarly, rule-based automation follows explicit, predefined instructions – if X happens, do Y.
Agentic AI, however, is about enabling true autonomy, moving past traditional approaches that include Human-In-The-Loop concepts. An AI agent is a program that can perceive its environment, reason about complex situations, autonomously plan its own multi-step actions to achieve a goal, and then execute those actions. Crucially, it can also adapt and learn from its experiences, even without definite instructions.
The core difference is this: traditional systems respond or follow rules, while agentic AI proactively acts with a goal in mind, iterating and adapting. It’s not just generating text; it’s independently taking steps to solve a problem or fulfill an objective, even interacting with external systems and tools as needed. This marks a significant shift from mere assistance to genuine autonomous action.
Beyond task automation, what complex, multi-step problems are agentic AIs uniquely positioned to solve that current AI models cannot?
Agentic AIs, unlike current Generative AI models focused on single-turn prompts or predefined workflows, are uniquely positioned to solve complex, multi-step problems by exhibiting genuine autonomy and goal-driven behavior.
Traditional Generative AI excels at specific tasks (e.g., generating text, creating code). Agentic AIs, as highlighted by Cloudflare’s advancements, go beyond by:
- Proactive Planning & Execution: They can break down complex objectives into sub-tasks, prioritize, and orchestrate actions across various tools and services without constant human intervention. Think of a complete travel booking from research to confirmation, adapting to real-time changes.
- Contextual Awareness & Memory: Agentic AIs maintain context across interactions, learning from past results and adapting their strategies. This “memory” allows for sophisticated, multi-stage problem-solving, unlike reactive, stateless models.
- Dynamic Adaptation & Error Recovery: If an unexpected obstacle arises (e.g., a flight is sold out), agentic AIs can dynamically re-plan, explore alternatives, and resume execution, a capability largely absent in rigid, current AI workflows.
- Human-in-the-Loop Integration: While autonomous, they can seamlessly incorporate human feedback and approvals at critical junctures, enhancing reliability and trust in complex processes.
In essence, agentic AIs tackle problems requiring continuous decision-making, adaptation, and multi-tool orchestration, transforming AI from a reactive assistant to a proactive, goal-oriented problem-solver.
How much autonomy should Agentic AI have in decision-making, and where should humans remain in the loop?
At Cloudflare, we champion a “Human-in-the-Loop” (HITL) approach for Agentic AI. While recognizing AI agents’ potential for autonomous decision-making and task execution, Cloudflare emphasizes that full autonomy isn’t always desirable or safe. We advocate for clear approval workflows where AI agents can request human authorization before executing sensitive actions.
This is crucial for high-stakes or ethically complex scenarios, preventing hallucinations or unintended consequences. Humans provide vital judgment, context, and oversight and are often better placed to understand the risks to business processes that fully Agentic AI may not.
Cloudflare’s platform and tools facilitate this balance, enabling agents to pause, seek human input, maintain state during review, and resume seamlessly once approved. This combination leverages AI’s efficiency with human expertise, ensuring responsible and effective AI deployment.
What infrastructure challenges (such as power, cooling, latency) arise when deploying Agentic AI at scale?
Deploying Agentic AI at scale presents significant infrastructure hurdles. Key among these are the demands on power and cooling. Agentic AI, especially with large language models, is incredibly compute-intensive, requiring more energy than traditional server racks. This translates to immense heat generation, necessitating advanced cooling solutions like direct-to-chip or immersion cooling to prevent overheating in data centers.
Latency is another critical concern. Agentic AI thrives on ultra-fast, low-latency communication between agents and across data centers globally. This necessitates a rethinking of network infrastructure, moving beyond traditional electronic switches to embrace technologies like photonic switching, which significantly reduce energy consumption and latency by routing data entirely optically. Cloudflare emphasizes running AI closer to users (at the edge) to minimize latency and ensure responsiveness.
Finally, managing state and persistent context for millions of simultaneous agent interactions is challenging. Cloudflare addresses this with “Durable Objects,” a serverless compute primitive that combines computation with storage, allowing agents to maintain context across interactions and scale efficiently without developers managing complex infrastructure.
What’s the next frontier for Agentic AI—will we see AI “agents” collaborating like human teams?
Cloudflare views agentic AI as the next frontier, shifting from AI that simply provides instructions to systems that autonomously execute tasks. This involves AI agents making decisions, learning, and adapting. While direct statements about “AI agents collaborating like human teams” can’t yet be explicitly detailed in a broad sense, our focus on the Model Context Protocol (MCP) and “human-in-the-loop” functionalities strongly hints at a collaborative future.
MCP allows agents to securely interact with external services and tools, facilitating complex multi-step workflows. Furthermore, Cloudflare emphasizes developers building agents that can persist context, remember past interactions, and even seek human approval for actions. This suggests a future where AI agents act as intelligent, independent collaborators within existing human-led processes, rather than fully autonomous, self-organizing teams of AI.
Will open-source models keep pace with proprietary Agentic AI systems, or will there be a widening gap?
Cloudflare sees open-source AI models rapidly closing the gap with proprietary Agentic AI systems, rather than a widening one. Thanks to innovations like Group Relative Policy Optimization, open-source models are now nearly matching proprietary performance at significantly lower costs. This democratizes AI development, enabling smaller organizations to deploy and specialize models.
While proprietary models offer immediate, robust solutions, open-source provides transparency, customization, and cost optimization. Cloudflare highlights a trend towards a commoditization of base AI capabilities, with open-source models becoming increasingly competitive. They emphasize that the “choice” between the two is a false dilemma, and hybrid approaches, leveraging the strengths of both, will likely become more common.
Cloudflare itself is actively supporting open-source AI development by providing tools and infrastructure to accelerate agentic AI building.
