What is a Coding Agent?

Where a code-completion tool suggests the next line as a developer types, a coding agent operates at the task level: given an instruction like "add pagination to this endpoint," it explores the relevant files, makes the edit, runs the test suite, reads the failure output, and corrects its own mistake — looping until the task is genuinely done or it hits a boundary it can't resolve alone. This requires more than language modeling; it requires tool use (a way to read/write files and execute commands), a feedback loop (tests, linters, type checkers) to verify its own work, and often a scoped environment — a sandbox or worktree — so an in-progress change can't corrupt a shared branch. The quality of a coding agent is measured less by how good its first draft is and more by whether it can recover when that first draft is wrong. As these agents get more capable, the human role shifts from writing code line-by-line to defining scope, reviewing diffs, and setting guardrails — the same shift that happened with build systems and compilers, one level up.