When Agents Test Other AgentsAugust 16, 2026 · 12 min readLi JinDora SSR developerI recently ran an experiment that sounds unnecessarily circular. Codex acted as a user: it proposed games, sent multi-turn requests, inspected the results, and asked for changes. Dora Agent acted as the developer. Across dozens of games, the work became more than a gallery. Codex summarized repeated failures and used them to improve Dora Agent itself. The games changed, and so did the software making them.
The Project as the Agent's External BrainAugust 16, 2026 · 5 min readLi JinDora SSR developerOne day I inspected Dora's application data and found an Agent database of roughly 255 MiB. That is not enormous by modern storage standards, but it was still growing. The diligent colleague had kept before-and-after file bodies for checkpoints, tool steps, parameters, results, and session data. Some orphaned tasks and checkpoints could also survive after their original session disappeared. The discovery changed the question from “How much can an Agent remember?” to “Which memory deserves which lifetime?”