Current project
Evolving review councils
When AI agents build software, a council of independent reviewer agents checks the work for bugs, security holes and risky changes. Here the council evolves: each generation tries to beat the best version so far, and a change is kept only if it beats the best version on a fixed set of hard tests, with held-out tests the agents never see to catch gaming.
Repeating the runs at different council sizes, on tasks of different complexity, shows where one more reviewer stops paying for itself.
97
Videos
Watch the experiments
Short explainers on how each experiment works, and what it finds once the runs are done.
Coming soon
Council calculator
A small tool for setting up a review council for your own task. It will estimate the council size, model mix and token cost, then write a prompt you can paste into any AI. The estimates will come from the experiment's results, so it will arrive after the runs.