Chat Is Out, Agents Are In: What Mollick's AI Tool Guide Says About 2026

Simon Willison flagged something worth sitting with: Ethan Mollick's guide to "which AI to use to do stuff" has changed shape in about a year, and the change tells you more about the industry than any single model release does.

A year ago the guide was a chatbot comparison: ChatGPT, Claude, Gemini, with o3, Claude 4 Opus, and Gemini 2.5 Pro as the top picks and Deep Research as the useful alternative mode. That's a chat-first framing — you pick a model, you have a conversation, you copy the output somewhere.

The shift: from "which chatbot" to "which agent"

Today's edition, per Willison, is organized around agentic systems — tools built to do "the equivalent of many hours of real human work in one go" rather than answer one prompt at a time. The named contenders are ChatGPT Work, ChatGPT Codex, Claude Cowork, and Claude Code. That's a real category shift, not a rebrand: the unit of work moved from a reply to a completed task.

The most striking absence is Gemini. Willison notes it has fallen off Mollick's list because Google still doesn't have an established entry in the Codex / ChatGPT Work / Cowork category — Gemini Spark exists but, in his words, "has yet to prove itself." For a company with Google's model and infra depth, that's a notable gap in the one dimension — agent packaging, not raw model quality — that's currently deciding developer mindshare.

The naming problem is not cosmetic

Willison also calls out something anyone evaluating these tools will hit immediately: the web-based modes and the desktop apps share names — ChatGPT Work, Claude Cowork — but not capability. The desktop apps are the more powerful implementation because they can be given access to your actual computer, not just a sandboxed container. On mobile, ChatGPT's "Work" mode is notably different from plain chat in one specific way: its Code Interpreter container gets internet access, which the default chat mode doesn't grant. Small detail, but it's exactly the kind of thing that silently changes what a task can accomplish and that a team evaluating tools for real engineering work needs to check rather than assume.

Why this matters if you're building with AI, not just chatting with it

This lines up with what we see day to day doing AI-assisted development: the interesting engineering question in 2026 isn't "which model scores higher on a benchmark," it's "which agent runtime, with what permissions, running where." A model embedded in a sandboxed web session behaves very differently from the same model with filesystem and network access on your own machine — different risk surface, different failure modes, different review discipline required before you let it touch production code or a live trading system.

That's also why "vibecoding" isn't a synonym for "turn off code review." An agent that can run for hours unsupervised needs the same guardrails a junior engineer would: scoped permissions, a clear diff to review, and a gate before anything ships. The tools have gotten more capable; the discipline around them has to scale with that capability, not lag behind it.

FAQ