An AI system that can do software work in multiple steps — reading a codebase, writing or editing files, running commands, and checking results — to complete a programming task. Unlike a basic autocomplete tool, a coding agent is meant to take action and iterate until it reaches a goal, often with some human approvals along the way.

For data reporters and newsroom developers, an AI coding agent can speed up the unglamorous parts of building and maintaining reporting pipelines. For example, it might scaffold a scraper for a local government site, write a script to extract tables from PDFs into a CSV, or refactor a messy notebook into a repeatable command-line tool — then add logging and basic tests so the work is easier to rerun and audit. The key is to treat the agent like a fast junior collaborator: give it a small, testable task, review diffs carefully, and run its changes in a sandboxed environment before they touch real data or production systems.

Coding agents are a specific application of AI agents: they plan, use tools, and can keep going over many steps. They often rely on reusable instructions (agent skills), can be split into parallel specialists (subagents), and may connect to outside systems through standards like Model Context Protocol. Because they can read and act on external inputs (like websites and documents), they also raise security concerns such as prompt injection and benefit from approval checkpoints (see plan mode).