A file-based manager and worker harness claims frontier-level coding from Qwen3.8 models

The authors describe "ledger-based zero-shot self-orchestration", a method that needs no training. Fresh instances of one model split a problem into tasks and coordinate through a shared workspace holding a plan, notes and the current solution. On the 100 latest hard LiveCodeBench problems, across eleven open and closed-weight models, they report gains of up to 25.6 percentage points on pinned backends. Orchestrated Qwen3.8 Flash Next reaches 93.0% pass@1 against 90.4% for Fable 5, at 9% of the cost. A locally served Qwen3.8-27B rises from 66.8% to 92.4%. The authors also write that "gains are not universal": some models were unchanged or worse, and they attribute the gain to decomposition and persistent context after reading transcripts. The README says the Opus 5 result came from an earlier scaffold, in a single pass through OpenRouter. Code is MIT-licensed and the paper and run data are CC BY 4.0.

The Hacker News thread had 165 points and 57 comments about seven hours after posting. The top objections are about the evidence. One commenter calls LiveCodeBench a saturated benchmark with likely contamination and notes there is no comparison against newer frontier models or on agentic work. Another points out the harnessed model was not compared with other harnessed models. A third says the run departs from the benchmark's own rules by adding a delegation loop. Several practitioners report using a similar plan, status and notes file pattern by hand with the same model.

Why it matters: The design is the useful part for anyone building factories: fresh-context workers, a manager, and written state on disk as the only shared memory. The headline claim is a different matter. It rests on one benchmark whose problems may overlap training data, and the paper's PDF was not read for this item, so treat the parity claim as the authors' unreplicated result. The repository does ship a runnable harness (--engine multiagent against --engine single), so a team can test it on its own tasks.