Verifiers v1: Train Agents Past the Context Window

On July 10, 2026, Prime Intellect shipped verifiers 0.2.0, a preview of the rewritten core they now call v1. It landed under the new verifiers.v1 namespace, and the announcement came with a sentence that got the RL crowd's attention: the legacy code path is frozen and will not be actively maintained. If you have an environment on the old API, it still runs. It just stopped being the future.

The reason this blew up over the last few days is not the version bump. It is what the rewrite makes possible. v1 tears an environment apart into three independent pieces — a taskset, a harness, and a runtime — so any task can be solved by any agent, anywhere. And because rollouts are stored as a directed acyclic graph of messages instead of a flat list, a single rollout that compacts its context four times yields four trainable samples. That is how you train an agent on a horizon longer than its own context window.