In plain terms
Autocomplete finishes the line a developer is typing. A coding agent takes a task, such as “fix this bug” or “add export to PDF”, and works like a developer who has just joined the team: it searches the repository, reads the relevant files, makes the change, runs the tests, reads the errors and tries again. What comes back is a finished change with a description, ready for review.
Why it matters
It moves the unit of delegation from a line of code to a task, and that changes how engineering teams spend their time: less typing, more specifying and reviewing. Routine work such as upgrades, test writing and small fixes can run in parallel. The bottleneck moves to review. An agent can produce plausible code that passes the tests it wrote itself and still misses the intent, and code that nobody on the team has read becomes a maintenance and security burden. Budget for review capacity, for automated tests and for usage costs, which grow with long runs.
Example
A software company gives a coding agent a ticket: invoice totals are off by one cent for some customers. In eleven minutes the agent finds the rounding function, reproduces the fault with a failing test, changes six lines, runs all 148 tests, sees them pass and opens a pull request that explains the cause. A developer reviews the change in ten minutes and merges it.
Most often confused with
Coding Agent vs. Code completion (autocomplete)
Code completion works inside the editor, one suggestion at a time, and the developer accepts or ignores each one. A coding agent works across the repository and the terminal, takes many steps on its own and hands back a complete change. Most current products offer both modes. The practical difference is where the person checks: every line as it appears, or the finished change at review.
Under the hood
A coding agent is a model in a loop with a small set of tools: search and read files, edit files, run shell commands and use version control. The harness adds a system prompt, project instruction files that describe conventions and commands, context management for long sessions, subagents for search or review, and permission modes that decide which commands run without asking. Agents run in a terminal, in an editor, or as cloud jobs that start from an issue and end in a pull request. Examples include Claude Code, OpenAI Codex, the agent features of GitHub Copilot, Cursor and Gemini CLI. Safeguards: a sandbox or container, no production credentials, restricted network access, protected branches, mandatory human review and continuous integration. Repository content and fetched web pages are untrusted input, so prompt injection applies here too. Public benchmarks such as SWE-bench and Terminal-Bench measure task completion; results on your own codebase matter more.