A conversational assistant: understand and prepare
It helps explain concepts, compare options, draft text or suggest code. Without access to your project, you provide the context and apply its suggestions yourself.
An agent: work in an environment
An agent combines AI with tools and a working loop: observe, act, check the result and adjust. With the necessary access, it can change files directly.
Claude Code and Codex
Claude Code by Anthropic and Codex by OpenAI are agentic development tools. They can inspect a project, edit code and run commands or tests. Their capabilities depend on the environment, available tools and permissions.
Example: a broken form
In a conversation, you share the code and error, then apply the suggested fix. An agent working in the project can locate the relevant files, prepare a fix, run available checks and present the changes for review. If checks fail, it can continue investigating.
The boundary is not absolute
Conversational assistants can also have web search, files and execution tools. Agents can simply answer questions. The distinction is the access provided and work delegated, not the chat window.
Which should you choose?
A conversation is often enough to learn, think or prepare content. An agent can help implement project changes. Define the expected result and review the work: passing tests does not guarantee that there are no errors.
Sources / Quellen: Claude Code · Codex CLI