Building long horizon agent worlds.

We think AGI is still missing two capabilities: long horizon planning and continual learning. Everything else seems to be arriving on schedule. Models today are superhuman at completing tasks and embarrassingly bad at making decisions, and we think the reason is simple; they never have to live with the consequences of their actions.

You can see this clearly in software. AI-generated codebases tend to go well for the first few development cycles and then get harder to work with, because the model made design decisions it never had to answer for (much like a short term contractor). For now this is mostly managed with careful human specs, and it hasn't improved much as benchmark scores have climbed.

Neither capability can develop in environments that end in hours. So we build simulated worlds that run for months, where agents act and the consequences of their actions come back to them. We're early, and there's a lot we don't know yet.

We don't think agents should be units of work that humans task like they do today. We think they will join the workforce, and we want to help them get there.

We're a small team from Google and Stanford. We think much of the path from here to AGI runs through worlds like ours, and we'll be sharing what we find inside them.