Prime Intellect releases Verifiers v1; DAG trace format eliminates quadratic data growth in agentic RL

Prime Intellect released Verifiers 0.2.0 on July 10, a ground-up redesign of its open-source RL environment library that solves two critical scaling problems for agentic training. First, it introduces a directed acyclic graph (DAG) message format that eliminates quadratic data growth—previously, each training turn stored a full copy of the prompt plus new completion, so trace size grew with the square of rollout length. Second, it decomposes the environment into three independently swappable layers (taskset, harness, runtime), replacing v0's monolithic design that required code rewrites whenever any component changed.

The v1 architecture splits responsibility: a taskset defines work (data, tools, rewards), a harness defines the solver (ReAct loop, Codex, Terminus 2, custom agent), and a runtime defines execution (local subprocess/Docker, or cloud sandboxes via Prime, Modal). This composability enables teams to run the same benchmark under different agents with only configuration changes—Terminal-Bench 2 was ported to v1 with a single small class. An interception server architecture multiplexes 32 rollouts per server and sits between agent and inference, enabling server-side reward hacking mitigation and tighter training-eval integration.

For practitioners: the DAG change shifts trace storage from O(n²) to O(n), unlocking training on trajectories longer than an agent's native context window—a structural blocker for long-horizon SWE and reasoning tasks. Prime Intellect's own testing showed zero overhead vs. Harbor running identical tasks. Architects building agentic RL systems should expect composable environment designs to become standard; monolithic environment coupling is now a liability. The library has 4,200+ GitHub stars and is integrated with Prime's hosting platform and vLLM.