Code Completion & Assistant
Continue, explain and revise code inside the editor
WHAT THIS CAPABILITY MEANS
Takes the code context around the cursor plus an intent — complete a line, change a function, explain the logic — and outputs an edit that can be adopted directly. Its constraint is consistency with the existing repository rather than producing an isolated snippet. Unlike code generation it makes local changes inside existing files and conventions, and therefore leans more on context.
How it is done
A common approach concatenates the current file, nearby symbols and retrieved related snippets into the prompt so the model completes or rewrites within full context; repository-level assistants use code retrieval to surface relevant definitions and call sites. Edit requests are often framed as generating a patch that is then applied, which makes review and rollback easy. The system is tightly coupled to the IDE and needs low-latency inference with incremental context updates.
Representative products
4GitHub Copilot
2021Completing and rewriting code in your editor from context
Cursor
2023A desktop code editor that treats the whole repository as context
Claude Code
2025A coding agent that carries out multi-step tasks from the terminal
Devin
2024An autonomous coding agent with its own shell, editor and browser
Organizations involved
Typical uses
- Line-by-line completion while coding
- Bug fixes and refactoring suggestions
- Explaining code and onboarding into an unfamiliar repo
- Writing tests and doc strings
How it is evaluated
- Suggestion acceptance rate
- Share of suggestions developers keep, closest to real experience
- pass@k
- One-shot success on tasks with tests
- Test pass rate for a single edit
- Whether the repository tests still pass after one edit
Limits and hard parts
- It edits the wrong variable or misses call sites, producing code that clashes with the surrounding semantics
- Repository understanding is capped by the context window, so distant constraints are missed in large projects
- It can emit outdated or insecure patterns that the author must review
Concepts behind it
Transformer Architecture
Replacing word-by-word relay with a room where everyone speaks at once, so long-range dependencies are one hop away
Retrieval-Augmented Generation
Rather than cramming knowledge into parameters, leave it outside and look it up on demand — an open-book exam instead of a closed-book one
Prompting & Alignment
Making a model helpful, honest and harmless is harder than simply making it bigger