AI Agents
AI agents are autonomous software systems that perceive their environment, reason about goals, and take independent action to accomplish tasks. Unlike traditional chatbots that respond to single prompts, agents operate in loops—observing, planning, acting, and learning—often for extended periods without human intervention.
The capabilities of AI agents have expanded exponentially. The autonomous task horizon—how long an agent can work independently on a complex task—has doubled from minutes to over 14.5 hours in just 18 months, according to METR benchmarks. This isn't incremental improvement; it's a qualitative shift in what software can do. An agent that can work autonomously for 14 hours can accomplish things that were previously impossible without a human team.
Modern AI agents are built on large language models but extend far beyond text generation. They can browse the web, write and execute code, interact with APIs, manage databases, coordinate with other agents, and use specialized tools through protocols like the Model Context Protocol (MCP). Frameworks like LangChain, CrewAI, AutoGen, the Claude Agent SDK, and the OpenAI Agents SDK provide the scaffolding for building these systems. At the platform level, agent operating systems like NVIDIA's OpenClaw are emerging to manage model routing, tool orchestration, memory, sub-agent coordination, and enterprise policy enforcement.
The Inference Explosion
Agents are the primary driver of the inference scaling phenomenon. Every user query that triggers an agentic workflow generates vastly more compute than a simple chat response. The agent reasons through chains of "thinking tokens," spawns sub-agents, queries tools, evaluates results, and iterates — a single user request might generate 100x more tokens internally than the final visible response. Jensen Huang at GTC 2026 cited this multiplier effect as the engine behind a computing demand increase of roughly one million times in two years, with inference growing 100,000x relative to training.
The economic implication: "Every SaaS company will become an Agent-as-a-Service company." The current generation of enterprise software — CRM, ERP, ITSM, HR platforms — will evolve from tools that humans operate to agents that humans supervise. This transition drives continuous background inference demand that dwarfs interactive usage.
The economic implications are the subject of what might be called the $211 billion question. In 2025, AI venture capital hit $211 billion—representing 50% of all global VC—yet only 6% of organizations report more than 5% earnings impact from AI. The outlier effect is dramatic: top-quartile AI users see 6x productivity improvements over the average. At Anthropic, merged pull requests per engineer increased 67% after introducing Claude Code. The adoption curve is doubling every four months.
This is the technology driving the transition from the Engineering Era to the Creator Era. Agentic engineering—using AI agents to build software—collapses the gap between a founder's vision and production code. Complex multiplayer platforms that once required months and large teams can be built by solo founders with agents over a weekend. The implications ripple across every industry: when intelligence becomes cheap and abundant, the bottleneck shifts from execution to imagination.