Case Study: Forge Across Many Sessions

This case study designs Forge for a task too large for one session, and the design makes specific promises. Any session can start knowing nothing and be useful within a few steps. No session's changes are kept unless the harness accepts them, so each session starts from the last accepted commit. The task is not reported as complete until a test that the agent could not skip has passed for every feature.

It makes no promise that the code will be well structured, that the feature list is right, or that a model working for 1,000 steps will avoid every error