Crabhaus

What it is

All the software wrapped around a language model to make it act: the system prompt, the tool catalog, the loop that executes tool calls and feeds results back, permission gates, and the rules for what stays in context. Anthropic's SWE-bench write-up defines an agent as exactly this pairing — a model plus its scaffolding.

How it works

Each cycle, the harness assembles everything the model sees: instructions, tool definitions, history, and the latest results. The model replies with an action; the harness parses it, runs the real tool, appends the observation, and calls the model again until a stop condition. Since history is re-sent every turn, harness design is dominated by managing that growing context.

A concrete example

Coding benchmarks score model+harness pairs, not bare models. Terminal-Bench submissions must name both an agent and a model, and Anthropic reported SWE-bench results "can vary significantly" from scaffolding changes alone — same weights, very different rank under two different harnesses.

Why it matters

Choosing between agent products means choosing between harnesses at least as much as models. A benchmark gap between two tools on the same model is scaffolding, not weights — and in your own stack, harness work is often a bigger lever than a model upgrade.