Disclosure: I maintain the open-source project Tura. This is not an independent review; I’m publishing the implementation and results so others can reproduce them and point out mistakes.
A long coding-agent task often repeats a familiar sequence: inspect the environment, update configuration, patch the bug, adjust tests, build, run tests and lint, then inspect Playwright screenshots. Not every transition requires a new model decision, but each tool call normally consumes another round.
Tura’s Macro layer batches deterministic steps into one tool call and returns structured results, including the exact failure point. If new evidence requires judgment, the macro stops and hands control back to the model; this is not blind batching.
On the same 60-task set we measured:
| Configuration | Passes | Pass rate | Observed tokens | Rounds | Estimated cost |
|---|---|---|---|---|---|
| Macro + backward reasoning | 48/60 | 80.0% | 229,695,477 | 2,017 | $221.138 |
| Macro Direct | 39/60 | 65.0% | 75,108,167 | 969 | $99.620 |
| Codex CLI Medium | 38/60 | 63.3% | 333,538,349 | 3,140 | $257.173 |
| Codex CLI High | 36/60 | 60.0% | 455,742,296 | 6,074 | $327.483 |
In this sample, Macro Direct achieved 1.7 percentage points more passes than Codex CLI Medium while using 77.5% fewer observed tokens and 69.1% fewer rounds. Against Codex CLI High, it used 83.5% fewer tokens and 84.0% fewer rounds, with a 5-point higher pass rate. Adding backward reasoning produced the highest pass rate, but also increased tokens and cost.
“Up to 84% fewer rounds” is not a universal guarantee. The useful denominator is total cost per completed, verified task, including failures, retries, and orchestration. These are aggregates from one 60-task set; other repositories can produce different ratios.
Code, scripts, tasks, and the reproduction method:
Where would you trust a deterministic macro in your agent workflow, and where should the model stop to inspect new evidence?
Top comments (0)