Disclosure: I maintain Tura.
A coding agent often spends a separate model turn on each part of a routine workflow: inspect the environment, edit package files, patch the implementation, update tests, build, run lint/tests, then inspect Playwright media.
Tura experiments with a simpler execution model: describe that sequence as a macro and let the runtime execute the steps together, while still returning structured results to the agent. The goal is to reduce orchestration turns without hiding failures.
On our current 60-task DeepSWE set:
| Configuration | Passes | Pass rate | Observed tokens | Rounds |
|---|---|---|---|---|
| Macro + backward reasoning | 48/60 | 80.0% | 229,695,477 | 2,017 |
| Macro Direct | 39/60 | 65.0% | 75,108,167 | 969 |
| Codex CLI Medium | 38/60 | 63.3% | 333,538,349 | 3,140 |
| Codex CLI High | 36/60 | 60.0% | 455,742,296 | 6,074 |
The implementation, task data, and benchmark methodology are public:
The important caveat is that fewer turns do not automatically mean lower end-to-end cost. Retries, failures, cache behavior, and completed-task rate all belong in the denominator. I would especially appreciate feedback on the benchmark design and cases where macro execution makes debugging worse rather than better.