Intel / article

SWE-chat v2: 230K real prompts to coding agents, openly released

Source: huggingface.co
Author: StanfordAILab
Date: Published Oct 2, 2026

Context

SWE-chat is a record of how developers actually work with AI coding agents, and its second version holds about three times as many sessions as the first. The dataset card lists 17,968 sessions and 229,909 user prompts in the October 2 release, up from 5,851 sessions and 62,544 prompts in the April version. Stanford AI Lab shared the release as a living dataset of prompts from real developers.

Each session goes well beyond a chat log. The data comes from public GitHub repositories whose developers used the Entire.io command-line tool to save checkpoints of their sessions with agents such as Claude Code, OpenAI Codex, Gemini CLI and Cursor. Because those checkpoints are tied to commits, the dataset connects what a developer asked, the 's thinking traces and tool calls, and the code that was finally committed, with a count of which lines the agent wrote and which a person wrote. That allows questions a coding benchmark cannot answer, such as how often developers correct or reject an agent's work, which the card tracks with a pushback label on user prompts.

The sample is narrow by construction. It includes only developers who used Entire's tool and kept those checkpoints in public repositories, so it describes that group rather than software development in general. The prompt-intent, pushback and session-success labels were assigned by a , and the line attribution comes from Entire's own records. Downloading the files also requires a Hugging Face login and agreement to the dataset's access conditions.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.

Read the source huggingface.co