"A Guide to Which AI to Use in the Agentic Era" is an article by Ethan Mollick, published on his One Useful Thing newsletter on 2026-02-17 (Source: Raw Sources/A Guide to Which AI to Use in the Agentic Era.md). It is the eighth guide to AI tools Mollick has written since the launch of ChatGPT, which he describes as marking a "very large break with the past" because AI use has shifted from chatbot conversations to agentic work. The article introduces a three-layer framework — Models, Apps, and Harnesses — for understanding AI products.
Framework: Models, Apps, Harnesses
Mollick distinguishes three layers of an AI product. Models are the underlying systems — GPT-5.2/5.3, Claude Opus 4.6, Gemini 3 Pro — and are what benchmarks measure and what labs race to improve. Apps are the products used to talk to a model, including websites, phone apps, and specialized tools such as Claude Code, Codex, and Cowork. Harnesses are systems that let an AI use tools, take actions, and complete multi-step tasks; a harness wraps a model in a set of capabilities such as web browsing, code execution, and file access, and the same model behaves differently depending on the harness.
The central argument is that the same model produces different results across products: Mollick describes asking "the same model, Claude Opus 4.6, asked the exact same question in three different apps and harnesses" and obtaining very different outputs. He concludes that apps and harnesses now matter more than model differences for most users.
The landscape as of February 2026
Mollick identifies three frontier models — Claude Opus 4.6, GPT-5.2 Thinking, and Gemini 3 Pro — which he describes as "remarkably close in overall capability." He advises selecting the advanced or thinking model manually, noting that defaults often use weaker versions, and states that free models are "optimized for chat, rather than accuracy" and much less capable than paid frontier models.
On coding tools, Mollick groups Claude Code, OpenAI Codex, and Google Antigravity, arguing they are not only for coders because "the ability to code is about being able to do anything that happens on a computer." He describes Claude Cowork as "Claude Code for non-technical work," running on the desktop, working with local files and the browser, and operating in a VM with security isolation. He also covers NotebookLM, Google's knowledge-base tool for making sense of documents, and OpenClaw, an open-source agent that ran locally and controlled a user's computer via WhatsApp or iMessage, which he calls "a serious security risk" but a "sign of where things are going."
Key claims
Mollick characterizes the move from chatbots to agents as "the most important change in how people use AI since ChatGPT launched." He reports having Claude Code build, test, launch, and deploy a complete e-commerce site — selling 80 PDF volumes of GPT-1 weights — in about an hour. He states that OpenAI and Anthropic have "clear leads over Google" in harness capability despite comparable model quality.
Provenance and related pages
The article documents the state of AI adoption and usage by knowledge workers, relevant to AI as Normal Technology in mapping the gap between what is possible and what most users do. The shift from chatbot to agent bears on AI Labor Disruption as a qualitative change in AI's displacement potential. Mollick's observation that the top models from Anthropic, OpenAI, and Google have converged in capability connects to the Fast-Follow Problem. The article also documents Claude Code and Claude Cowork as agentic products from Anthropic.