An AI agent that reads, edits and runs code on its own to get programming jobs done.
A coding agent is an AI agent for programming. On its own, it finds and reads files, edits code, runs it to test it and fixes what breaks, over and over. Claude Code, Codex and Cursor's agent are well-known examples.
In 2022, GitHub Copilot suggested code but couldn't run it. On SWE-bench, a 2023 test built from 2,294 real GitHub issues, the best model then solved just 1.96%, and the test became the yardstick for coding agents.
Claude Code became generally available in May 2025, and in February 2026 OpenAI said Codex had written a whole internal product of about a million lines. Still, when it launched Codex, OpenAI told people to always review agent-written code themselves.
Think of it like this
Picture two friends helping with maths homework: one just reads out answers, the other works each problem, checks it and redoes it if it's wrong. Code autocomplete is the first friend; a coding agent is the second. But it's still your homework, so the final check is yours.
For example
Say "The photos on our club website won't load, fix it," and it finds the right files, fixes the cause and reopens the page to make sure it works.
Common mix-ups
Mix-upAnything a coding agent does can always be undone.
ActuallyClaude Code saves files before editing them so you can roll back, but its official documentation says actions on outside systems, such as databases or deployments, can't be undone that way. Cursor's checkpoints also restore files only. Set the agent to ask permission before it runs commands like that.
Mix-upA high benchmark score means it will handle my requests just as well.
ActuallyThe August 2026 RealSWE study found that 88% of real user requests contained only a description of the problem, while only 7% of benchmark tasks took that form. Rewriting the tasks the way real users write cut solve rates by 6.4 percentage points on average. Spelling out the behaviour you want, and why, helped a lot.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- When a coding agent says "All fixed!", how can you check?
- Activity
- In groups, write a website-fix request two ways: a one-line request in everyday words, and one that spells out the behaviour you want and why. Attach a test list to each to decide when the job is really done. Then read the SWE-bench (E0273) and RealSWE (E1075) records and compare benchmark tasks with real requests.
Records about this
Source-checked moments from AI history.
A strand that breaks down the gap between benchmark scores and what users experience in the field into "a difference in input distribution" rather than "a difference in ability". The finding that writing down the desired
An internal product of about 1 million lines written entirely by Codex; over 5 months, 3–7 engineers merged about 1,500 PRs
Company-reported: SWE-bench Verified 72.5% for Opus 4 and 72.7% for Sonnet 4, Terminal-bench 43.2% for Opus 4. Priced like the previous generation ($15/$75 and $3/$15 per 1M tokens), with measures for higher AI Safety Le
Later became the de facto objective function for coding-agent development