Autocomplete vs chat vs agent
There are three distinct generations of AI tooling, and conflating them sets the wrong expectations:
- Autocomplete (in-IDE completion) suggests the next line at the cursor. Context: the open file and its surroundings. You decide; the tool only types.
- Chat — question and answer in a separate window. Context: the code you paste. You copy the result back into the project yourself.
- Agent — works inside the project: reads files, edits them, runs commands, looks at the result and corrects itself. Context: the repository itself.
The key difference is the loop. The agent turns "read → change → run → inspect → fix" on its own. That is why the most important element of an agent task is the loop's stopping condition (the build passed, tests are green).
| Property | Autocomplete | Chat | Agent |
|---|---|---|---|
| Context source | The open file | Text you paste | The repo, command output, tools |
| Modifies files? | At line level | No | Yes, across many files |
| Runs commands? | No | No | Yes (with permission) |
| Main risk | Accepting without looking | Context-free answers | Sprawling unreviewed changes |
| Fits | Typing boilerplate | Learning a concept | Multi-step tasks, refactors, debugging |
The agent's power is also its risk: it changes files and runs commands on your behalf. That is why every serious agent tool has a permission system — not an extra obstacle, but the tool's security model.
📚 Sources and documentation
- Claude Code overviewofficialcode.claude.com
- Subagentsofficialcode.claude.com
Splitting a task across parallel subagents, each with its own context.