Tura
Tura: 16.7% better performance, 77.5% fewer rounds.
Start new thread

Tura - Macro execution for fewer coding-agent turns

by•
New 60-task benchmark and reproducible scripts: Tura groups deterministic repository work—environment checks, patches, builds, tests, lint and media inspection—into reusable macros. Macro Direct passed 39/60 tasks in 969 rounds versus Codex CLI Medium at 38/60 in 3,140 rounds. Adding backward reasoning reached 48/60. Open source, with the dataset, logs and methodology published.

Add a comment

Replies

Best
Maker
📌
Hi Product Hunt — I maintain Tura. This update is about evidence, not just a token-saving claim. We ran the same 60 long-horizon tasks across Macro Direct, Macro + backward reasoning, Codex CLI Medium and High, counting failures, retries, total rounds and observed tokens. The result that surprised me: Direct nearly matched Medium's pass rate (39 vs 38) with 969 vs 3,140 rounds, while backward reasoning raised the pass count to 48 but cost more. Repo and benchmark scripts are public. I'd especially value feedback on whether cost per successful task is the right denominator.

The benchmark page is honestly a nice touch, basically showing your work instead of just waving around vague claims about token savings. Love that it's local and open source too.

Maker

Thanks — that was exactly the goal. The benchmark is only useful if someone else can inspect the tasks, rerun it, and point out where the accounting is wrong.

honestly been wanting a coding agent that actually reads the repo before touching it, so tura scratching that itch. the open source angle is a big plus too.

Maker

That repo-first behavior is the part I care about most too. The macro layer only helps after Tura has mapped the project and knows what evidence to check.

Congrats on the launch, love that it's open source and local. One thing that would help me trust it more is a built-in diff preview mode where I have to approve each change before it hits my repo. Right now even a careful agent can surprise me with edits I didn't catch in the logs.