Testing shows coding harnesses can be harmful or hugely helpful depending on the model
JoshPurtell · x · 2026-11-07
Josh Purtell tested pi, swe-mini, and claude code harnesses against gpt-6-luna, inkling 2 small, and inkling small. Counterintuitive finding: non-codex harnesses were neutral or harmful for luna and inkling small, but massively helpful for inkling 2 small. Whether an external coding harness pays off depends heavily on the specific model — worth benchmarking per model before wiring one in.
More from coding & agent
- Loop engineering: hand specs to the agent, humans only review the goal — dotey · 2026-11-07
- duckdb-graphrag: GraphRAG with two tables and recursive CTEs, no graph DB — JeremyCMorgan · 2026-11-07
- Factory's coding agent Droid lands in Microsoft Teams: assign tasks via @mention — matanSF · 2026-11-07
- CodeRabbit podcast: ship less, not more, in the AI era — nutlope · 2026-11-07
- 16 MCP servers, 35 parallel agents — and it still ships slop — weswinder · 2026-11-07
- Dev ships a database-less MCP server that answers from your backup ZIP — itsnoddydev · 2026-11-07