Tura Collapses Agent Tool Calls Into One Macro Tool, Reporting 35.8% Fewer Turns Than Codex CLI

GitHub / Hacker News·low signalXBlueskyLinkedInCopy link

Tura (Tura-AI/tura, AGPL-3.0, ~570 stars) is an agent runtime that replaces the usual pile of small tools with a single `command_run` macro tool, letting the model build a multi-step execution tree in one LLM turn instead of round-tripping per call. Self-reported benchmarks across 270 sessions on 20 DeepSWE v1.1 tasks claim a balanced config at 35.8% fewer turns and 31.1% fewer tokens than Codex CLI with an 80% success rate versus 63.3%, and a direct config at 69.1% fewer turns and 77.5% fewer tokens at 65% success. Marked low because the numbers are vendor-authored on a 20-task sample with no independent replication, but the architectural bet — fewer, fatter tools — is testable in an afternoon.