AI Coding Assistants
AI that completes, writes and fixes code, from GitHub Copilot's preview in June 2021 to agents that resolve real software issues; in 2025, 84% of developers used or planned to use such tools.
Open in the interactive tree →GitHub Copilot, built on OpenAI's Codex model, was announced on 29 June 2021 and sold from June 2022. In February 2025 it gained an agent mode that runs commands, and in May 2025 a coding agent that opens draft pull requests for review. SWE-bench (October 2023) turned real GitHub issues into a test, and the best model then fixed 1.96% of the tasks. Code can be compiled and tested automatically, which makes it easier to train and score than most work.
As of now
The Stack Overflow survey of 2025 found 84% of developers using or planning to use AI tools (76% in 2024) and 47.1% using them daily, yet 45.7% distrust their accuracy and 66% name answers that are 'almost right, but not quite' as the top frustration. On SWE-bench Verified the leading published score was 80.9% in February 2026, after which OpenAI said it would stop quoting it. METR's randomized trial (July 2025) found 16 experienced developers took 19% longer with AI on their own repositories; its follow-up (February 2026, 57 developers) pointed to speedups of about 18% for returning and 4% for new developers, but METR called the data unreliable because many developers declined to work without AI.
Where today's AI could help
High AI leverage. Code can be compiled and tested automatically, so output is cheap to check and models can be trained and scored on pass/fail results; limits are vague specifications, large codebases and review time.
- Generate tests and fixes and check them by running the test suite
- Resolve real bug reports in open-source projects
- Review and refactor large codebases
- Write the training environments for the next generation of coding models
Shown so far
Prerequisites
- Programming Languages1957They write and read code in the languages people already use.
- Open Source & Linux1991Models learn from the huge body of public open-source code.
- Pretrained Language Models2018Copilot started as a language model trained further on public code.
- Large Language Models2022Today's assistants are large language models used through chat and tools.
Unlocks
- AI Agents2025Coding was the first area where agents worked on long tasks; tests check the result.