An AI coding assistant is a developer tool that uses a large language model to help write, complete, explain, debug, or refactor source code, typically integrated into a code editor, IDE, or command-line interface. Early versions offered line-by-line autocomplete; by the mid-2020s the category expanded to include agentic tools that can read a whole codebase, plan multi-file changes, run tests, and iterate with limited supervision, a capability enabled by tool use.
History
The category traces to statistical and rule-based autocomplete in older IDEs, but the modern era began with GitHub Copilot, announced by GitHub and OpenAI in June 2021 and built on a code-tuned descendant of GPT-3. Copilot's 2022 public release popularized inline, ghost-text suggestions trained on public code repositories, and its underlying Codex model was evaluated on the HumanEval benchmark OpenAI introduced alongside it. Competitors followed quickly, including Amazon's CodeWhisperer, Tabnine, and Replit's Ghostwriter. A second wave shifted from autocomplete to full agentic editors and command-line tools: Cursor (2023), an editor forked from VS Code with codebase-aware chat and multi-file editing, and Claude Code from Anthropic (2025), a command-line agent with direct file-system and shell access. By 2025, benchmarks such as SWE-bench had become a standard way to measure how well these agents resolve real GitHub issues on their own.
How it works
Most assistants combine a code-specialized model with editor-level context: the open file, cursor position, related files, and sometimes a whole repository indexed for retrieval. Agentic variants add tool use, letting the model read files, run a linter or test suite, execute shell commands, or query documentation, then use the results to revise its next step, a loop that applies retrieval-augmented generation and step-by-step reasoning to a live development environment. Some products expose this through the Model Context Protocol, letting an assistant connect to external tools and data sources through a standardized interface.
Impact and criticism
Studies of Copilot users, including internal GitHub research and academic follow-ups, reported productivity gains on routine tasks, with some randomized studies finding developers completing tasks over 50% faster, though effects vary widely by task type and experience level. Critics have raised concerns about code quality and security, since models can reproduce insecure patterns learned from training data; about intellectual property, since suggestions have sometimes closely reproduced licensed code, feeding into broader AI copyright disputes; and about skill atrophy, the worry that reliance on generated code weakens programmers' independent debugging and design ability. Proponents counter that assistants free developers from boilerplate and let them focus on architecture and review. By 2025 the debate had extended into the more radical practice of vibe coding, where a user directs an assistant almost entirely in natural language with little manual code review.