Open-source macro layer slashes coding-agent turns by 68% on 60-task benchmark

Tura, an open-source macro execution layer, reduces the number of model turns needed for coding agents by batching routine steps like editing, testing, and building. On a 60-task DeepSWE benchmark, the macro with backward reasoning achieved an 80% pass rate using 2,017 rounds, compared to Codex CLI High's 60% pass rate with 6,074 rounds. The project is public on GitHub, with caveats that fewer turns don't guarantee lower costs due to retries and failures.