Field Note: Claude.ai, Claude Code, Cowork: Which. When. Why?

There's a question floating around that deserves a straight answer: what are these tools actually for, and why did Anthropic build them in this order?

The answer isn't about product categories. It's about what emerged from real usage.

March 2023: Claude.ai

Conversational interface. Accessible. The baseline. People could talk to Claude, get responses, iterate. No barrier to entry.

February 2025: Claude Code CLI

Anthropic built something powerful for software engineers doing software engineering work. A terminal tool. Agentic. Could read your codebase, write code, run commands, all from the command line. Natural entry point for that work.

But something else happened. Knowledge workers, people doing research, synthesis, analysis, complex problem-solving observed the power in Claude Code. They saw what it could do. And they started using it for knowledge work. They had to be tenacious. The CLI wasn't designed for them. It was unfamiliar. But they persevered, and it worked.

September 2025: Claude Code VS Code Extension

In parallel, Anthropic released a graphical interface for Claude Code. VS Code integration. Same agentic capability as the CLI, but with a slightly easier entry point. For software engineers, some adopted the extension exclusively. Some stayed with the CLI. Some used both depending on the task.

But knowledge workers saw it too. Some who'd struggled with the CLI found the extension more approachable. Some stayed with the CLI and got comfortable with it. Some moved between them depending on the task. Broader cohort of knowledge workers could now access the same capability, just through different doors.

But it was still a software engineering tool. Knowledge workers were using engineering interfaces to do knowledge work.

January 2026: Cowork

Anthropic was watching. They saw software engineers moving fluidly between CLI and VS Code extension. They saw knowledge workers fighting through software engineering tools to do knowledge work those tools were never designed for. They realized something: there's a whole category of work. Research, synthesis, analysis, complex thinking that needs agentic capability. But it doesn't need to just live in a terminal or an IDE.

Cowork isn't Claude Code rebadged. It's the same agentic capability, but with a fundamentally different interface, constraining but optimized for the knowledge worker experience. Streamlined, task-oriented, designed for people who don't need raw terminal or IDE complexity, and who want to leverage scheduled, autonomous task execution.

But not everyone moved. Some knowledge workers who'd gotten comfortable with Claude Code CLI or the VS Code extension stayed there. Some moved to Cowork. Some use both depending on the task. The tools are now options, not gatekeepers.

In practice, from my own work

Claude.ai is where I do conversational work, stateless, lightweight Q and A. I use Projects extensively for context-rich research.

Cowork runs my recurring tasks. Scheduled work, hourly jobs. It's my agentic runtime for automation.

Claude Code in VS Code is where the technical work happens, things like building MCP Servers. I typically lean toward the extension, but I oscillate between that and the CLI depending on what feels right for the task. Multiple agents, raw access, native feel. I use it most because it's the engineering tool I used as an engineer, and I still use today.

They're all doing the same fundamental thing, accessing tools, directories, executing, but the interface changes which work feels natural. That matters more than you'd think.

The pattern

But here's what's interesting: this isn't unique to me. People naturally gravitate toward whichever tool fits their particular blend of technical skill and knowledge work needs. Some lean entirely toward, or go all in on Claude.ai, Claude Code and / or Claude Cowork. Some, like me, oscillate across all three depending on the work. There's no "right" adoption path. The system works because it offers multiple entry points into the same capability. You'll find your own equilibrium.

One thing worth noticing: sometimes the most important product insights come from watching audiences you didn't expect use your tools in ways you didn't anticipate. Knowledge workers weren't the target audience for Claude Code. But they showed up anyway, found value, and that unexpected adoption signal is what shaped Cowork. We're seeing this pattern again, Mythos wasn't designed for cybersecurity, yet it has proven to be a powerful defense and attack tool, more on that here.

Last thoughts are that when unexpected audiences discover value, I think that's often signal worth listening to.