Every hour your team spends rewiring evaluation infrastructure around a new agent architecture is an hour not spent improving the model itself.

Bundled environments have been quietly killing training velocity

The standard approach to agentic RL forced data, agent logic, and runtime infrastructure into a single coupled package. Swapping one piece meant rebuilding the whole thing, which made scaling experiments slow and compositional evaluation nearly impossible.

Three decoupled pieces now do what one brittle bundle used to

Prime Intellect Releases Verifiers v1 splits that bundle into three composable layers: a taskset that defines the data, tools, and scoring; a harness that solves the task and produces a rollout; and a runtime that executes the rollout either locally or inside a sandbox. A managed interception server sits between the agent and the inference endpoint, recording traces, setting sampling parameters, and rewriting tool responses to suppress reward hacks during training. Each server multiplexes up to 32 concurrent rollouts by default, and a pool scales elastically with observed load.

Infrastructure engineers feel this gap the most, but they are not alone

• ML engineers running large-scale RL training who need faithful token-level trace recording without hand-rolling proxy logic
• Research engineers evaluating coding agents with tools, compaction, and subagents who currently rebuild harnesses for each new benchmark
• AI platform teams who need to swap runtimes between local dev and sandboxed production without touching task or agent code

The separation of concerns is the whole point: any taskset runs under any compatible harness, which means experiments that previously required a rewrite now require a config change.

The market for agentic training infrastructure is moving faster than most teams can track

OpenAI, Google DeepMind, and a growing tier of open-source labs are all publishing agentic benchmarks that assume multi-step, tool-using agents as the baseline evaluation target. If your evaluation stack still assumes single-turn rollouts, you are already measuring the wrong thing, and the gap between what you can evaluate and what your agents can do will only widen.

What you can build with it today

• Run any taskset under a ReAct loop, a CLI agent, or a custom harness
• Record full inference traces for token-in reinforcement learning training
• Rewrite tool responses mid-rollout to block reward hacking during training
• Scale concurrent rollouts elastically without changing task or agent code

Pricing not listed — check our directory.

The composability is real, but the v1 namespace signals this is still early

Shipped under the verifiers.v1 namespace inside release 0.2.0, this is a preview of a rewritten core, not a production-hardened GA release. Teams adopting it now should expect the API surface to shift.

Alternatives worth knowing

EleutherAI’s lm-evaluation-harness covers breadth across static benchmarks but was not built for multi-step agentic rollouts. If your requirement is sandboxed execution with tool interception rather than benchmark coverage, the architectural fit differs significantly.

Agentic evaluation infrastructure is becoming its own engineering discipline

The teams shipping reliable agentic models in 2025 are the ones treating evaluation infrastructure as a first-class engineering problem, not an afterthought. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.