A coding agent is an AI-based system that uses development context to select and carry out work toward a software development objective.
This operational definition applies the AI agent concept to development. The implementation determines which steps the system can choose and execute, which require human input, and which remain unavailable. The label alone does not specify an IDE, CLI, model, protocol, or level of autonomy.
A suggestion and a delegated step differ
A generated patch is proposed content. Applying it is an operation performed by a person or an execution mechanism. A coding agent may participate in selecting and carrying out such steps, but an interface that only displays suggested text does not establish delegated execution of that text.
Describe the actual work rather than infer it from the assistant's name. Inspecting source, changing an artifact, selecting a test, and requesting work in another system are different possible operations. No single one is required in every coding agent, and permission for one does not automatically cover the others.
Development work can execute more than an edit
A development task can involve repository or dependency code as well as the agent's own proposed changes. For example, npm's version 11 scripts documentation describes lifecycle stages around test execution and scripts run through a shell. That is a property of the selected tooling, not a requirement that every coding agent use npm or a shell.
When such tooling is available, review what executes and under which identity. The submitted diff does not necessarily describe accesses or effects produced while preparing it. An execution environment can expose resources beyond the files the task asks to edit.
Describe the delegated scope
Specify the development objective, relevant repository or artifacts, permitted operations, execution environment, and decisions retained by people. A task to prepare a change is not, by itself, authorization to publish it or alter a shared service. The available credential may permit more than the assigned task.
A tool call is one way to request an operation. It does not establish the complete work behind that interface or the outcome of its execution. Keep the requested step, applicable permission, and observed result distinct when describing what the coding agent did.
The enterprise coding-agent rollout guide explains how to turn an authorized workflow into a bounded pilot with explicit acceptance criteria.
The coding agent governance pillar develops the authority, environment, and evidence requirements behind that delegation.