ThinkingBox releases database-outcome training for AI agents

The October 22, 2026 announcement introduces ThinkingBox-Training for developers training agents that change database records. At the end of an episode, checks compare the resulting state with the required outcome: wrong, missing or extra side effects fail. Most checks follow fixed rules; a minority use a model judge for narrow requirements. The public sample exercises the interface; it is not the private training pool behind the reported experiment.

The preparation guide provides a loader adapter: code that turns task identifiers into runnable cases. Keeping it outside the repository makes dataset location and task selection runtime inputs. Its public example contains 39 task identifiers, all outside the benchmark list; developers must validate every one before using the list, then can connect their own authorized tasks. The preparation command records counts, selected identifiers, hashes, statistics and a run manifest outside the repository.

The evaluation guide separates saved training state from a model ready to serve: distributed checkpoint parts must first be merged. The merged weights support inference, not resuming training. For the documented 507-task test with 20 repetitions, the evaluation script checks that every expected task-and-repetition pair appears exactly once, without system-error rows, before aggregating results. If infrastructure errors occur, the guide directs developers to rerun only missing or system-error cases and withhold a final score until coverage is complete.