Multi-model coding team draws interest

A coding workflow that splits a problem across several AI roles drew 389 points and 192 comments on Hacker News. Memdoor’s setup uses a planner, two ideators, a solver, a manager and a repair worker. Each starts with a fresh context and passes information through files, while a script, not an AI model, decides whether the answer passes the tests.

The workflow’s test results cover ten LiveCodeBench Hard problems, run five times each. Memdoor reports an average of 9.8 problems passed, compared with 9.6 for the authors’ harness using the same model and 8.0 for Fable 5 in a single call. The report notes that the sample is small and says 9.8 and 9.6 are the same number. Fable 5 wasn’t run in Memdoor’s harness, and the source describes the claim as parity with the authors’ harness, not a gain over it. So the claim that this workflow beats Fable 5 isn’t independently verified in the available report.

For an operator, the useful idea isn’t that a team of agents has proved better than one model. It’s that this gives you a concrete way to test whether dividing work into roles helps on your own tasks, with a script checking the output. That kind of repeatable comparison is what we build into an AI command centre.