A Guide to Which AI to Use in the Agentic Era

The piece argues that the shift from chatbot use to agent use changes how people should choose AI. The model still matters, but the app and the harness now shape what the same model can actually do, from chatting to research, coding, spreadsheets, and other multi-step work.

Reading notes

#

• “Using AI” used to mean back-and-forth chat, but now it can mean assigning tasks to an agent that uses tools.
• Choosing AI now depends on three things: models, apps, and harnesses.
• Models are the underlying systems that determine reasoning, writing, coding, image work, and benchmark performance.
• Apps are the products used to talk to models, such as the main websites and phone apps for ChatGPT, Claude, and Gemini.
• Harnesses are the systems that let a model use tools, take actions, and complete multi-step tasks.
• The same model can behave very differently depending on the harness it runs in.
• For advanced use, the author says the paid plans matter because free models are optimized for chat, not accuracy.
• The main frontier models named are Claude Opus 4.6, Gemini 3 Pro, and ChatGPT 5.2 Thinking.
• For serious work, the author says to select the smarter model manually rather than relying on default auto mode.
• ChatGPT, Claude, and Gemini all include chatbots with different bundled features such as research, study modes, image tools, and data connections.
• Claude.ai and ChatGPT are presented as stronger than Gemini on harnesses for work like writing code, producing files, and doing research.
• Claude Code, OpenAI Codex, and Google Antigravity are described as coding tools that can access a codebase, terminal, and testing loop.
• Claude for Excel and PowerPoint are presented as specialized harnesses that let the model act like a junior analyst inside those programs.
• Claude Cowork is described as Claude Code for non-technical work, running on the desktop with local files and browser access.
• NotebookLM is presented as a tool for making sense of large amounts of material and turning sources into slides, mind maps, videos, and podcasts.
• OpenClaw is described as a local agent that can browse, manage files, send emails, and run commands, but the author warns it is a serious security risk.
• The closing advice is to start with one major system, pay for the advanced model if needed, and use it on real work rather than demos.