Agentic AI

I get a version of the same question at almost every conference and customer meeting I attend: "Agentic AI sounds exciting, but what does it actually mean for how we operate our infrastructure day to day?" It's a fair question. The term gets used so broadly that it can feel more like marketing than a meaningful shift in how cities work. So let me answer it directly: with a real platform, real architecture, and real results.

Together with our partners at AIREV, Intel evaluated the OnDemand Agentic AI platform on Intel® Core™ Ultra Series 3 processors. What we found is not just a faster dashboard. It's a fundamentally different way of operating critical infrastructure, one where your systems understand intent, take action, and protect themselves, all without waiting for a human to read an alert.

What "Agentic" Actually Means in Practice

Most smart-city deployments today are reactive. Sensors generate data. Dashboards display it. Operators make decisions. That model puts humans in the loop for every single event — and cities generate thousands of events every second. You simply cannot staff your way to the level of responsiveness that modern infrastructure demands.

Agentic AI changes the operating model entirely. An AI agent doesn't just observe, it reasons, plans, and acts. It can be given a goal (keep water quality within safe parameters, optimize energy consumption during peak hours, ensure air quality alerts reach residents within 30 seconds) and it will execute the actions necessary to achieve that goal autonomously, continuously, and across connected systems simultaneously.

Think of it less like a smarter dashboard and more like a knowledgeable colleague who is always on duty, never misses an anomaly, and can execute a remediation workflow faster than any human operator.

The Four Layers You Need to Understand

The OnDemand platform is organized into four layers, and understanding each one helps you see exactly where Agentic AI plugs into your existing operations — and where it takes over.

Layer 1: Physical Infrastructure. This is your starting point. The buildings, streetlights, wind turbines, solar panels, water pumps, and environmental sensors (CO₂, CO, ammonia, TDS, humidity) that already exist in your city. You don't need to rip and replace anything. OnDemand integrates through standard MQTT publish-subscribe protocols, which means it speaks the language your sensors already use.

Layer 2: Real-Time Connectivity. The MQTT backbone provides bidirectional control and persistent telemetry storage. This is the nervous system; it ingests data from every connected asset in real time and stores historical records that fuel the intelligence layers above it. Bidirectional control is the critical distinction here: it's not just listening, it's also capable of issuing commands back to actuators instantly.

Layer 3: Agentic AI Core. This is where the intelligence lives. Natural-language processing, tool orchestration, predictive analytics, and conditional automation logic all operate here. This layer is what allows an operator to type "Is the water pump running?" and receive an accurate, real-time answer — without navigating a menu or opening a specific application. More importantly, it's what allows the system to automatically shut down a pump when water TDS levels exceed a safety threshold, send an SMS alert to maintenance staff, and log the event — all within seconds, and all without human initiation.

Layer 4: Human Interaction. Operators remain in control through conversational commands, role-based dashboards, automated reports, and alerts. The key design principle here is governance without friction: humans set policy and receive escalations, while agents handle execution. This matters enormously for public-sector accountability, you always have an audit trail of what the agent decided and why.

Three Ways to Deploy Agentic AI Today

If you're asking how to start, here's how I recommend thinking about it:

Start with environmental monitoring. OnDemand tracks CO₂, CO, ammonia, toluene, acetone, and alcohol in real time. You can configure agents to fire alerts, adjust HVAC systems, or trigger emergency notifications the moment a hazardous reading appears. This is one of the fastest deployments to stand up and immediately demonstrates life-safety value.

Move to energy optimization. Solar-generation tracking, device-runtime analytics, and intelligent load balancing are natural extensions once your telemetry layer is live. Agents can disable streetlights automatically when solar output drops below a defined threshold not because a rule said, "turn off lights at X watts," but because the agent understands the relationship between generation capacity and load demand and acts accordingly.

Add predictive maintenance as your confidence grows. This is where Agentic AI delivers its highest long-term ROI. By analyzing historical sensor trends and runtime data, the platform identifies anomalies that precede failure and schedules maintenance during optimal windows before anything breaks. The shift from reactive maintenance to predictive maintenance alone can dramatically reduce downtime and extend asset lifespan across water, energy, and climate-control infrastructure.

What the Performance Numbers Tell Us

One of the most common concerns I hear from infrastructure operators is latency. When an acetone sensor fires a hazardous reading, you need action in seconds not minutes. Our evaluation on the Dell XPS 16 powered by Intel® Core™ Ultra X7 358H (Panther Lake) using the Qwen3 8B language model at INT4 quantization showed that the GPU path delivers a Time to First Token of just 125 ms and a Time to 10 Tokens of 538 ms. That is real-time performance for an LLM-driven safety decision executing at the edge.

The NPU path, while slightly slower on initialization (TTFT of ~1.2 seconds), delivers lower inter-token latency at 32.2 ms/token compared to the GPU's 41.3 ms/token, meaning smoother, more consistent token streaming during sustained reasoning tasks. For city infrastructure where workload patterns vary, running GPU for fast-response alerts and NPU for steady-state analytics gives you the best of both worlds on a single edge device.

The practical implication: you do not need a cloud round-trip for agentic decisions. The intelligence can live at the edge, on commercially available hardware, with latency profiles that meet operational safety requirements.

Preserving Human Oversight at Scale

I want to address one concern directly, because it comes up in every conversation about autonomous systems in public infrastructure: What about accountability?

The answer is architectural. Agentic AI in OnDemand is designed around human-in-the-loop governance for policy, with agent-in-the-loop execution for operations. Your operators define thresholds, safety protocols, and escalation rules. The agents execute within those boundaries and every action is logged, time-stamped, and available for audit. When a situation exceeds the agent's defined authority, it escalates to a human and waits. You set the boundaries; the agent respects them.

This is the architecture that makes autonomous infrastructure politically viable, not just technically possible. Cities can move confidently toward cognitive operations because the governance model scales with the technology.

The Shift That Matters

The move from smart city to cognitive city isn't about replacing operators, it's about elevating what operators do. Instead of monitoring dashboards for anomalies, your team focuses on policy, resilience strategy, and capital planning. The agents handle the execution layer that used to consume most of the available attention.

If you've been waiting for Agentic AI to mature enough for production deployment in critical infrastructure, the answer from our evaluation is clear: it's ready. The platform works, the hardware is available today, and the architecture is designed for the governance requirements of public-sector operations.

The question is no longer whether agentic AI belongs in your city. It's how quickly you can move from reactive to cognitive.  Learn more in our white paper. 

Hassnaa Moustafa is a Principal Engineer at Intel focused on Edge AI and critical infrastructure. She is a co-author of the Intel × AIREV OnDemand smart city solution paper.