Code agent
Also known as: coding agent · software engineering agent
A code agent writes, runs, and fixes code: it reads a repo, edits files, executes, reads errors, and iterates.
- The loop is: understand → edit → run → observe → repeat.
- Sandboxed execution is essential.
- The most mature and widely used agent category today.
Why it leads: code is the one environment with *precise, fast, verifiable* feedback — tests pass or fail. That tight feedback loop is what makes code agents reliable enough to ship.