The AI landscape has shifted from chatbots to agents—here's how to choose what to use

Ethan Mollick, an AI educator and researcher, published a framework in February 2026 arguing that the practical use of AI has fundamentally shifted. Until recently, AI use meant back-and-forth chatbot conversations. Over recent months, agentic systems that can autonomously execute multi-step tasks have become practical, requiring users to reconsider how they evaluate and select AI tools.

Mollick proposes dividing the AI landscape into three interconnected layers: models, apps, and harnesses. Models are the underlying intelligence systems—Claude Opus 4.6, GPT-5.2, and Gemini 3 Pro represent the current frontier—determining reasoning ability, writing quality, coding capability, and multimodal performance. Apps are the interfaces through which users access models: chatgpt.com, claude.ai, and gemini.google.com, along with specialized tools like Claude Code and OpenAI Codex. Harnesses are the systems enabling models to use tools, take actions, and complete autonomous workflows without human intervention at each step.

The same model produces markedly different outcomes depending on its operating environment. Claude Opus 4.6 accessed via claude.ai's standard interface retrieves current information and cites sources; the same model inside Claude Code receives a virtual computer, code terminal, and web browser, enabling it to research, build, and test projects autonomously. This distinction has major practical implications: GPT-5.2 answering a question differs substantially from GPT-5.2 Thinking navigating websites and assembling presentations.

Current frontier models have converged in overall capability and generally produce fewer errors than previous generations. However, serious use of advanced AI requires at least a $20 monthly subscription to access top-tier models and applications; free models remain optimized for conversational speed rather than accuracy. Mollick emphasizes that the default models provided in many interfaces often use weaker variants—GPT-5.2 defaults to an 'auto' mode that frequently selects less powerful sub-models—requiring manual selection of advanced options like GPT-5.2 Thinking Extended or Claude Opus 4.6 with extended thinking for complex tasks.

The emergence of agentic systems has complicated the question of which AI tool to use. Previously, model selection dominated the decision. Now, app design and harness capabilities matter equally, making the choice context-dependent on the specific task and desired outcome.