SWE-Agent setup and GitHub issue demos with GPT-4

This tutorial, uploaded April 5, 2024, covers the early release of SWE-Agent, an open source coding assistant for fixing GitHub issues. No numbered software version is supplied. The presenter cites a 12.29% SWE-bench result with GPT-4 and compares it with Devin; those benchmark figures do not establish the chance of solving any particular issue.

The walkthrough uses Docker, Miniconda and the repository's environment file, then builds a Docker image. The presenter encounters an unresolved setup error on an Apple silicon Mac and moves to Lightning AI to continue. He configures a required GitHub token and an OpenAI API key for the demonstrated run. These are the setup steps shown in the 2024 source, rather than current installation guidance.

SWE-Agent's custom commands let the model search a repository, inspect files in 100-line views and edit code. A linter rejects syntactically invalid edits. In the first demo, the agent attempts to fix an issue in its own repository but stops after exceeding the configured $2 cost limit, without a confirmed resolution. An author's second demo reproduces a SymPy matrix insertion bug, edits the indexing logic and produces a patch that the author says passes the relevant tests.

The demonstrated AI agent uses a hosted model API. The presenter wants local LLM support but says he cannot find a straightforward way to use it at that time; the tutorial does not demonstrate running a model locally.