Agentic AI is not just about running a single model. It’s a whole system challenge. The tough part lies in building an environment that supports many moving pieces working together.
Dynatrace recently launched new AI agents designed to reveal the hardest part of AI operations. Their agents focus on key issues like CPU capacity, memory management, and data access. They also emphasize policy-aware tool use and system observability.
Intel conducted thousands of experiments on agentic AI workloads. From these tests, they shared five lessons for enterprise leaders. One stands out: plan capacity based on agents per virtual CPU, not just the number of agents. This shifts focus from quantity to efficient resource use.
Monitoring agent task latency, especially the 95th percentile, proves more useful than looking at average CPU usage. This helps spot bottlenecks that might hide in average numbers. Intel advises defaulting to scale-out systems for hosting agents, saving scale-up for workloads needing heavy compute or special architectures.
Why Agentic AI Is a Systems Problem
Agentic AI depends on the full system working well. It’s not just about inference or individual model output. The system has to handle task orchestration, data access, running tools, managing latency, enforcing governance, and scaling infrastructure. All these parts matter.
Most existing agentic AI solutions don’t measure overall system performance. They focus instead on isolated tasks. This leaves gaps in understanding real-world, enterprise-level performance. Metrics like task success rate, cost per task, time per task, throughput, agent density, and latency become crucial.
Intel extended Terminal-Bench to evaluate AI agents better. This tool now includes profiling, telemetry, and replay capabilities. It helps teams understand how agents perform in complex workloads, not just isolated tests.
Industry Moves and Innovations
NVIDIA expanded its Agent Toolkit in July 2026. The new additions include PhysicsNeMo and CUDA-X libraries. These tools help build autonomous AI engineers with physics skills, accelerated solvers, and quantum chemistry capabilities.
NVIDIA’s PhysicsNeMo library trains AI models with physics knowledge. CUDA-X libraries accelerate solving complex problems in engineering and chip design. For example, cuISS speeds up large sparse linear systems in simulations. CuDSS focuses on electronic design automation and scientific simulations. CuEST supports high-accuracy quantum chemistry simulations.
NVIDIA’s Nemotron 3 Ultra leads in agentic RTL coding on Verilog design benchmarks. This model can run locally or on premises, giving companies more control and privacy over their data.
Several big names use NVIDIA’s tools. Cadence integrates Nemotron and CUDA-X in its AI Super Agent to speed chip design. Synopsys builds secure, accelerated agentic workflows with NVIDIA’s models. Siemens uses NeMo Gym, Nemotron, and CUDA-X for multi-tool workflows in its Fuse EDA AI Agent. Samsung applies CUDA-X and cuLitho for computational lithography and thermal-stress analysis.
ChipAgents uses NVIDIA’s toolkit to create AI agents specialized in chip design and verification. Silvaco scaled a massive 3.2-billion-mesh-node optical simulation in under four hours using NVIDIA GPUs. Keysight accelerated electromagnetic simulations by up to 10 times with NVIDIA cuDSS. Samsung, Synopsys, and TSMC integrate NVIDIA cuEST for quantum-chemistry workloads, achieving speedups between 10 and 50 times.
Timothy Costa from NVIDIA said, “Engineering has reached an inflection point. AI can now work with tools of physics, simulation and design.” NVIDIA adds, “With NVIDIA Agent Toolkit, developers can build agentic engineers that reason using physics, run complex simulations and generate high-fidelity data to become a new engine for innovation in chip and system design.”
Challenges and the Road Ahead
Experts agree that agentic AI needs a reliable environment to succeed. Cost management, control, and trust are key to moving from pilot projects to full production. Human oversight remains vital. Sean Michael Kerner notes that humans in the loop build trust and help agentic systems learn and remember.
Meanwhile, companies like Anthropic explore new frontiers with products like Claude Science. Google DeepMind worries about scaling millions of interacting agents and the complexity this brings. Startups claim breakthroughs in fixing bottlenecks that limit large language models.
On the enterprise side, Gartner named XMPro as a sample vendor for Agent Orchestration in their 2026 Hype Cycle for AI in IT Operations. This shows growing industry focus on managing these complex AI systems effectively.
The path to agentic AI at scale is clear. Plan capacity by agent density per vCPU, monitor task latency closely, and scale out systems by default. This approach helps businesses build AI that works reliably, controls costs, and inspires confidence as they grow.
Based on
- Dynatrace’s new agents can reveal the single hardest part of AI operations — thenewstack.io
- Building the enterprise environment for agentic AI | MIT Technology Review — technologyreview.com
- A single AI agent conversation can look perfect and still be broken, leaders from LangChain, Conviva and CoreWeave said at VB Transform 2026 | VentureBeat — venturebeat.com
- XMPro Named as a Sample Vendor for Agent Orchestration category in the Gartner® Hype Cycle™ for AI in IT Operations 2026 | AP News — apnews.com
- NVIDIA Expands NVIDIA Agent Toolkit With NVIDIA PhysicsNeMo and CUDA-X Libraries to Transform How the World Engineers, Designs and Builds | Markets Insider — markets.businessinsider.com
