Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents -- A Source-Code Study of Eleven Systems

A coding agent depends on software around its model: the loop that runs tools, manages working context and enforces permissions. A source-code study of eleven coding systems reports that the inspected runtimes used custom execution loops and code-search tools rather than general-purpose agent frameworks or vector-based code retrieval, which searches through numerical representations of meaning.

That finding gives builders concrete architecture patterns to examine. It does not establish that those choices outperform every alternative: the authors inspected implementations rather than running all eleven systems on a shared task set. Their evidence describes how these agents were assembled, not a ranking of their effectiveness.