A source study of eleven coding harnesses finds none use an agent framework or embedding retrieval

The authors read the source of eleven coding harnesses, including Claude Code, Codex CLI, Gemini CLI, OpenHands, Aider, Hermes and OpenCode, plus Databricks' Omnigent as a contrast case. They describe an 83-page paper with 13 cross-cutting observations and 29 recurring design patterns, and do not benchmark or rank the systems. Across roughly four million lines of code, they report that no runtime imports a general-purpose agent framework such as LangChain, LangGraph or AutoGen, and none retrieves code with vector embeddings; the field uses hand-rolled async loops plus ripgrep, tree-sitter and glob. SKILL.md skills lead MCP in adoption, 9 of 11 systems against 8 of 11.

One observation concerns loops: Mini-SWE-Agent's minimal linear loop reports results in the same range as OpenHands' event-sourced engine, so loop sophistication does not predict benchmark performance. The paper also source-diffs eight systems across one quarter and reports that Codex adopted Claude Code's hook vocabulary verbatim. It closes with 18 design recommendations and a 90-line minimal harness. These are the authors' findings from their own reading; this edition read the abstract, introduction and the loop sections, not the whole paper, and nothing here has been independently reproduced.

Why it matters: If you are choosing a stack, the reported pattern is that production harnesses sit on plain loops and deterministic search, with extension through skills and MCP. That is evidence about what shipped systems do, not about what works best.