Campfire runs Jev and Claude Code, Codex on the same task for comparison
Ahead of a LangChain session on building a harness for Jev, the author recommends the open-source Campfire tool, which runs multiple coding backends in parallel on the same repo and prompt for self-built evals.
LangChain is running a session next week on how to build a harness with Jev. A harness. For a model. That's the real signal. When you're routing decisions through a fast, cheap classification layer in an agent loop, the question isn't "does Jev work?" It's "does Jev work better than what I'm already running on my specific data and eval set?" The only way to answer that is to run them side by side. Same task. Same context. Different models. That's what Campfire is for. Campfire is an open-source tool that runs multiple AI coding backends on the same engineering task in parallel - Claude Code, Codex, Goose, Gemini CLI, whatever you have configured. Same repo, same prompt, same constraints. Outputs side by side. When Jev ships a coding backend, you can drop it in and see exactly how it stacks up against your current stack on real work. Not vendor benchmarks. Your own eval set. The harness for the agentic dev workflow already exists. http://github.com/stretchcloud/campfire https://x.com/LangChain/status/2101077173604143503
This post quotes another post: “LangChain livestream covers building a harness with Jev”