Not a single feature — a composition. You start with a language model and attach the pieces it needs to solve one specific problem. Think of them as primitives you snap together.
A persistent workspace bundling related chats, shared files, and instructions so context carries across conversations.
Details Claude retains about you and your work, so it doesn’t start from zero each time.
The standing directions that shape how the agent behaves for its particular job.
A link into an outside service — Drive, Slack, a database — so the agent can read and act on real data. The user-facing on/off switch for one integration.
Model Context Protocol — the open standard connectors are built on, so one integration works across many AI apps. The plumbing under a connector.
A packaged set of instructions and resources that teaches the model a specific job — build a slide deck, fill a PDF — loaded only when relevant.
Discrete actions the model can invoke: search the web, run code, read a file. The verbs available to the agent.
The scripts and logic that wire everything together — orchestrating when the model thinks, when it calls a tool, and what happens with the result.
What the assembled agent produces: a document, a chart, a working app. The output you keep. (This page is one.)
The chat app. You converse, it answers, makes artifacts, uses connectors. Where most people meet the agent.
A desktop runtime for knowledge work — research, sheets, slides, docs — where the agent runs tasks for you rather than just chatting.
The runtime for developers. The agent edits files, runs commands, and builds across a whole codebase from your terminal or editor.