The same model can fail in one product and finish the job in another. The difference is the harness: which tools it may call, what it is allowed to remember, how the context is cut when the window fills, and who checks the result before the next step. Swapping in a larger model is one change. Changing that surrounding system is another, and it can matter as much.
We file an agent when we can say what it did, what it was not allowed to touch, and who was still in the loop. A waitlist and a screen recording are not enough. Odyssey-3 is on this shelf because the claim is a policy trained on a learned world, not because the preview looks like a game.
A chat launch and an agent launch can share a lab and still be different news. GPT-6 moved the Chat tab. Work and Codex stayed put. That split is the kind of fact this shelf exists to keep visible.