Skip to content

A Guide to Which AI to Use in the Agentic Era

Key Ideas

  • Choosing an AI now requires thinking about three layers: Models (the underlying AI brains like GPT-5.2, Claude Opus 4.6, Gemini 3 Pro), Apps (the products you interact with, such as chatgpt.com or claude.ai), and Harnesses (the systems that let AI use tools and complete multi-step tasks autonomously).
  • The same model behaves very differently depending on its harness -- Claude Opus 4.6 in a chat window is a fundamentally different experience from Claude Opus 4.6 operating inside Claude Code for hours autonomously.
  • For serious work, always manually select the advanced/thinking model rather than relying on default "auto" mode, which often routes to weaker models.
  • Claude Code, OpenAI Codex, and Google Antigravity represent the most powerful agentic harnesses, primarily aimed at coders but capable of broad autonomous work.
  • Claude Cowork is a desktop agent for non-technical work that can operate on local files and browser, running in a secure VM -- described as "Claude Code for non-coders."
  • NotebookLM remains a standout tool for making sense of large document collections, with citation-linked answers and AI-generated podcasts.
  • The shift from chatbot to agent is characterized as the most important change in how people use AI since ChatGPT launched.

Summary

Ethan Mollick argues that the question "which AI should I use?" has become significantly harder because AI usage has evolved beyond simple chatbot conversations into agentic workflows. He introduces a three-layer framework: Models (the AI brains), Apps (the interfaces), and Harnesses (the tool-using infrastructure that lets models do real work). The top frontier models -- GPT-5.2, Claude Opus 4.6, and Gemini 3 Pro -- are now close enough in raw capability that the app and harness matter more than model choice for most users.

The article surveys the current landscape of AI apps and harnesses. The chatbot interfaces from the big three providers have diverged in their bundled features: Gemini offers strong image/video generation tools, ChatGPT bundles Deep Research and shopping, while Claude focuses on Deep Research and agentic capabilities. For doing real work, OpenAI and Anthropic have clear leads over Google in harness capability, with both able to write code, produce files, and do extensive research.

Beyond chatbots, Mollick highlights Claude Code and Claude Cowork as particularly transformative. He describes having Claude Code autonomously build an entire print-on-demand website in about an hour with minimal human involvement. For newcomers, he recommends picking one system, paying the $20/month, selecting the advanced model, and using it for real work rather than demos. For experienced users, he suggests trying NotebookLM (free) and then the Anthropic ecosystem of Claude Code and Cowork.

Relevance to Economics Research

This guide is highly relevant for economics researchers navigating the rapidly expanding AI tool landscape. The models-apps-harnesses framework helps researchers understand why the same model can produce very different results in different contexts -- critical when choosing tools for tasks like data analysis, literature review, or code development. The emphasis on agentic tools like Claude Code and Cowork points toward a future where researchers can delegate multi-step research workflows (data collection, cleaning, analysis, visualization) to AI agents rather than managing each step manually through chat.