Agentic AI
Agentic AI refers to AI systems that act autonomously to accomplish multi-step goals — not merely answering a single query but planning, using tools, accessing external data, executing code, and coordinating with other AI instances (sub-agents). The key distinction from earlier conversational AI is goal-directed action across an extended time horizon.
Anatomy of an Agentic System
jensen-huang describes an agentic system as having four core capabilities (Lex Fridman Podcast #494, 2026):
- Access to ground truth — reading file systems, databases, and knowledge stores
- Research capability — querying the internet, running computations, checking facts
- Tool use — interacting with APIs, software, hardware
- Sub-agent spawning — delegating sub-tasks to additional AI instances, creating “large teams” of AI workers
Huang’s analogy: a humanoid robot that enters a home and uses the existing tools (microwave, scalpel, hammer) rather than transforming its own hand into each tool. The robot reads the microwave manual on first encounter (via internet access) and becomes instantly expert. This mirrors exactly how agentic AI systems function.
The Agentic Inflection (2025–2026)
openclaw (an agentic AI platform launched ~2025–2026) is described by Huang as the “iPhone of tokens” — the fastest-growing application in history, doing for agentic systems what ChatGPT did for generative systems. nvidia anticipated this architectural pattern two years before its public emergence, designing the Vera Rubin rack (with fast storage accelerators and the Vera CPU) specifically for agentic workloads that bang on file systems and tools. See rack-scale-computing.
Agentic Scaling as the Fourth Scaling Law
Agentic AI enables the fourth of Huang’s four ai-scaling-laws: spawning sub-agents multiplies throughput without architectural changes to the base model. Sub-agent outputs become new training data, feeding a flywheel: pre-training → post-training → test-time reasoning → agentic experience → new synthetic data → back to pre-training.
Security Considerations
NVIDIA’s security framework for enterprise agentic deployments (NemoClaw, OpenShell) enforces a “two out of three” rule: an agent may have any combination of (1) access to sensitive data, (2) code execution, (3) external communications — but never all three simultaneously. This limits blast radius from a compromised or misaligned agent.
Impact on Employment and Programming
Huang predicts the population of “programmers” grows from ~30 million to ~1 billion as “coding” becomes specification in natural language. The radiologist analogy: despite computer vision exceeding human diagnostic accuracy since ~2019–2020, the number of radiologists increased because AI throughput expanded the total volume of scans, requiring more human oversight and specialist interpretation. Similarly, software engineers will grow in number as AI automates mechanical coding but creates new problem-solving demand. See artificial-general-intelligence for Huang’s AGI framing.