Feynman ships open-source research audit agent

The real differentiator is not “another agent framework,” it’s the verification layer: Feynman is trying to make research outputs harder to bluff. That makes it more interesting than generic deep-research wrappers, especially for people who care about provenance, reproducibility, and paper-vs-code mismatch.

• –`feynman audit <arxiv-id>` is the sharpest feature here because it targets a real pain point: papers that sound right but don’t line up with the code.
• –The workflow is opinionated in a useful way: researcher, reviewer, writer, and verifier are separated instead of letting one agent improvise everything.
• –It’s broader than a paper reader; the project also supports literature reviews, simulated peer review, and experiment replication on local or cloud GPUs via Modal and RunPod.
• –The “skills only” install matters because it makes the research stack portable into other agent environments, not just the standalone CLI.
• –At roughly 2.4k GitHub stars and MIT licensed, it already has real traction for an early open-source project, but it still reads like an ambitious v0 with a clear thesis.