Raven Wants to Be the Harness That Manages All Your Other Agent Harnesses

EverMind-AI has published Raven, an open-source project it describes as a harness of harnesses, aimed at developers juggling more than one agent framework at a time. The repository went up on GitHub and was submitted to Hacker News under a Show HN post that had drawn 9 points and zero comments as of this writing. The name and framing point to a coordination layer rather than yet another single-purpose agent runner.

The pitch is narrow but pointed: agent harnesses — the scaffolding that feeds a model its tools, context and loop control — have multiplied, and each one tends to assume it owns the whole session. Raven's stated purpose is to sit above them and manage several at once. That matters for anyone running benchmarks, mixing providers, or comparing how different harnesses handle the same task without rewriting glue code for each.

The timing tracks with how the agent tooling market has split over the past year. OpenAI's Codex CLI, Anthropic's Claude Code, and a long tail of open projects each ship their own harness conventions, and switching between them means reimplementing the same plumbing. A meta-harness promises a common control plane, though the hard part is that harnesses differ precisely in the details that resist abstraction: retry behavior, permission prompts, and how they expose tool output.

There is little public detail to evaluate yet. The Hacker News submission had no comments, so no third party has weighed in on how Raven handles the awkward cases, and the repository link is the only primary source. Nine points is modest traction — enough to surface on the new page, far short of the hundreds that notable Show HN launches accumulate.

The project also faces a naming collision risk and a scope question. Harnesses that try to unify other harnesses tend to either stay thin and lose value or grow thick and become just another harness to integrate against. Whether Raven threads that needle depends on how much of each underlying tool's semantics it preserves rather than flattens.

What to watch is straightforward: whether EverMind-AI publishes documentation on which harnesses Raven supports today, and whether developers running multi-harness evaluations show up in the comments to say it solves a real problem. Until then, Raven is a claim with a repository attached, and the burden is on the code to show the abstraction holds.