Agent Harnesses Are a Bigger Variable Than the Model

A paper called HarnessTax directly measured how much the agent scaffolding around a model affects coding performance, separate from the model itself. The finding is that the harness matters a lot, and that lighter harnesses (the Pi harness was mentioned) can cut costs in half compared to Claude Code or OpenAI Codex's harnesses while maintaining competitive results. The thread pushed back on terminology (harness vs. agent) but the underlying point held.

This connects directly to the Anthropic/OpenClaw thread, where Anthropic told users they could no longer use third-party harnesses like OpenClaw with their Claude Code subscription. The stated reason was that these tools put outsized strain on systems. The actual reason, commenters suspected, is that efficient third-party harnesses use tokens more effectively and cost Anthropic more per subscription dollar than the official client.

The pattern is that the agent harness market is diverging from the model market. OpenSpec, also on HN today, is a lightweight framework for specifying agent behavior. Builders are actively shopping for better scaffolding independent of which model they use, and the providers are starting to notice.