What Are AI Coding Agents?
Learn how agentic coding tools differ from autocomplete and what makes them capable of multi-step, multi-file work.
From autocomplete to autonomy
Early AI coding tools suggested the next few lines as you typed. A coding agent is different: given a task in plain language, it can read multiple files, form a plan, make edits across the codebase, run the build or test suite, read the failure output, and try again — with little or no per-step guidance. The unit of interaction shifts from a keystroke to a task.
The core loop: plan, act, observe
Most coding agents run a loop of reading context, deciding on an action such as editing a file or running a command, executing it through a tool, and reading the result before deciding the next action. This loop is what lets an agent recover from a failing test or a typo without a person intervening after every step, but it also means a bad early assumption can compound across many actions if nothing checks it.
Tools the agent can call
An agent is only as useful as the tools it is given: reading and writing files, running shell commands, searching the codebase, browsing documentation, or calling a linter. The scope of those tools defines the agent's blast radius. A read-only research agent is low-risk; an agent with unrestricted shell access can modify or delete anything the underlying user account can touch, which is why permission prompts and sandboxing matter.
Context and codebase awareness
A coding agent's usefulness depends heavily on how well it can find the relevant files, existing conventions, and related code before writing anything — the same codebase-awareness problem that separates good and bad autocomplete tools, but at a larger scale because an agent may need to explore dozens of files to complete one task.
Where humans still need to be in the loop
Agents are good at mechanical, verifiable, multi-step work: refactors, fixing a well-described bug, writing tests for existing code, updating a dependency across many call sites. They are weaker at judgment calls — architectural trade-offs, ambiguous requirements, and anything where 'looks correct' and 'is correct' diverge, such as security-sensitive code or subtle business logic. Review agent output the way you would review a capable but unfamiliar contributor's pull request.
Practical exercise
Give a coding agent a small, well-scoped task in a real repository — such as adding tests for one existing function or fixing a specific, reproducible bug — and watch which files it reads before editing, whether it runs the test suite, and how it responds when a command fails. Note one thing it did well and one place you would not trust it unsupervised.
Sources and further reading
These primary or specialist references informed the concepts in this guide. Product details can change, so verify current documentation before implementation.