What's this about?
An essay proposing that a general-purpose agent harness should be built from Unix primitives — files, processes, pipes — arguing LLMs are already "Unix-native" and the OS itself should be the agent's scaffold, with a VM as the sandbox-that-is-the-harness.
What exactly is a "harness"?
The harness is everything around the model that turns it into an agent: prompt construction, state, tool invocation, execution loops. Claude Code, Codex CLI, and Grok Build are all harnesses around different models. The joke in the thread — "are we all going rock climbing?" — reflects how fast the jargon spread; an arxiv paper this same week ("Harness Handbook") argues harness quality now rivals model quality in determining capability.
What's the community saying?
HN 173 points, 87 comments
- Generic vs domain-specific — the sharpest thread: several builders report domain-specific harnesses already beat generic ones, undermining the essay's premise.
- "This already exists" — critics say the proposal re-invents agent sandboxes and Claude Code/Codex themselves; the Unix-mapping is elegant but adds little.
- Determinism camp — a popular counter-philosophy: agents should be *as deterministic as possible* — more code, less model.
- The open question — how much do labs post-train models on their own harness's formats? If a lot, a truly generic harness may never compete.