Definition
An AI agent that can inspect a software project, change it through development tools, evaluate the result, and iterate toward a software outcome. Its working loop usually combines a model with repository access, file editing, shell or IDE tools, test feedback, and rules about when to continue, stop, or ask for help.
The label does not tell you how much authority the agent has. One coding agent may only prepare a diff in an isolated workspace. Another may open a pull request, respond to review, merge, deploy, or operate production systems. A useful description therefore includes the authority profile as well as the generation capability.
What makes it an agent
The system chooses and sequences intermediate actions in response to what it observes. It may search for relevant code, form a plan, make an edit, run a test, diagnose the failure, and revise the patch. Real-session studies show why the loop needs external evidence: an agent's confident completion claim can diverge from what it actually changed or verified.
Distinguish it from nearby terms
A code generator can return source text without seeing the project or running it. A coding assistant keeps a human at the center of the interaction. A coding agent owns a multi-step execution loop, although a human may still steer or approve that loop. A software factory is the larger production system in which one or more coding agents may operate.
Check your understanding
Two products both advertise a "coding agent." One can edit a temporary worktree and return a patch. The other can merge to the default branch. Are they operationally equivalent? No. Before comparing them, document repository access, network access, credentials, allowed tools, approval gates, and the external state each can change.