AgentV-RL turns verifiers into agents

This is less a product launch than a useful signal: reward models are starting to look like agents, because single-pass scoring is too brittle for hard reasoning.

• –Forward and backward verification directly targets a common failure mode: plausible final answers hiding broken intermediate logic.
• –Tool use matters because reward models without external grounding struggle on math, code, and knowledge-heavy tasks where "sounds right" is not enough.
• –The practical bet is distillation: use expensive multi-agent verification to train a smaller deployable verifier, then spend inference budget only where it improves selection.
• –The tradeoff is compute and complexity, so developers should read this as a direction for high-stakes evals and test-time search, not a drop-in scoring API.