Batching multi-step coding agent tasks into one turn to cut tokens
When using like Codex or Claude Code, a typical workflow needs a separate turn (one round-trip call to the model) for each step: editing a config file, fixing the bug, editing the test script, building, running tests and lint, and — if Playwright is involved — reading the rendered output. That can add up to six separate turns. Instead, bundling all these steps into a single plan (listed as step1, step2, and so on) and sending it as one reportedly cuts the number of turns by 40-80%, with a similar reduction in token usage.
The author tested this on Codex using GPT-5.6 sol high and published a benchmark on a (Tura-AI/tura), evaluated on DeepSWE-style full-repository rewrite tasks. One , called "Macro + backward reasoning," reportedly passed 48 out of 60 tasks (80%), used about 230 million tokens across 2,017 rounds, and cost roughly $221.
Key points
- Typical flow needs up to six separate turns: edit config file, fix bug, edit test script, build, run tests/lint, and (for Playwright) review rendered output
- Proposed approach bundles the steps into a single structured as step1 through step5
- Author claims 40-80% fewer turns and a similar reduction in token usage
- Benchmark published on GitHub (Tura-AI/tura), tested on DeepSWE-style full-repository rewrite tasks
- One reported ('Macro + backward reasoning') passed 48/60 tasks (80%), used ~230M tokens over 2,017 rounds, at an estimated cost of about $221