Sign InOpen Brain
AI EngineerVideoSource Linked

Reinforcement Learning without Verifiable Rewards — Will Brown, Prime Intellect

For agent work without deterministic answers, mine production traces into tasks, build controllable simulators, and use scaled judges to derive rubrics while watching for reward hacking.

AI Engineer · Jul 31, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Deterministic rewards work for math answers, tests, linters, and known database states, but reports, purchases, and customer interactions lack one clean answer. The proposed loop uses **production traces**, **controllable simulators**, and **scaled judges** to construct training signal.

Practical Implication

Treat deployed-agent traces as the emerging task distribution, then mine failures and calibrate task difficulty. In simulators, plant a known end state and work backward; use multiple judges and retrospective analysis to extract reusable rubrics before training.

Agent-Ready Context
Deterministic rewards work for math answers, tests, linters, and known database states, but reports, purchases, and customer interactions lack one clean answer. The proposed loop uses **production traces**, **controllable simulators**, and **scaled judges** to construct training signal.

Treat deployed-agent traces as the emerging task distribution, then mine failures and calibrate task difficulty. In simulators, plant a known end state and work backward; use multiple judges and retrospective analysis to extract reusable rubrics before training.

Proxy rewards remain exploitable at their boundaries, and real deployments may contain information that reinforcement learning never discovers. The talk therefore adds supervised signal from the environment, while acknowledging that environment and reward design still require substantial manual work today.
Connected Context · Feed7 Judgment

This extends agent training beyond domains with exact rewards by turning production traces, backward-built simulations, multiple judges, and retrospective rubrics into provisional supervision. It confirms simulation as a bridge from deployment failures to training tasks, but keeps the boundary clear: proxy rewards remain exploitable, simulations may omit real information, and substantial human environment design is still required.

From Agent Traces to Agent Simulations — Rustem Feyzkhanov, Snorkel AITrace replay provides the fixed evaluation environments that this training loop would need before production failures can become comparable tasks and reusable reward signals.SimulationMaxxing: How we ship agents 20× faster — Aman Gupta (Nubank) + Shreya Rajpal (Snowglobe)Nubank’s workflow reinforces simulation as an iteration multiplier while supporting this Signal’s caution that simulated outcomes require calibration against production data and human review.The Prime Intellect Stack — Will Brown, Prime IntellectThe composable environment stack is an implementation substrate for the proposed trace, simulation, evaluation, and reinforcement-learning loop without requiring agent harness rewrites.Harness Engineering is not Enough: Why Software Factories Fail — Dex Horthy, HumanLayerThe proposed scaled-judge training loop does not remove the maintainability and ownership risks identified here; human review remains a necessary boundary when proxy rewards cannot establish production quality.
Context Map
agentcodingresearch#harness-engineering#agent-evals#agent-reliability
Uncertainty
Proxy rewards remain exploitable at their boundaries, and real deployments may contain information that reinforcement learning never discovers. The talk therefore adds supervised signal from the environment, while acknowledging that environment and reward design still require substantial manual work today.