Intelligent Hillclimbing (judgmentlabs.ai)

Most teams treat evaluations (evals) as a checklist rather than a lived feedback loop, so they end up “climbing the wrong hills.” The post argues that generic LLM judges and one-off, human-written rubrics often fail to generalize across domains and quickly go stale as user behavior shifts, producing false confidence and slower iteration. Because evaluation defines what engineers optimize, poor eval design leads to misaligned improvements in agent behavior. The big takeaway: successful evals must be grounded in production signals — real interactions, corrections, and preferences — not crafted abstractions or static benchmark sets. The proposed fix is an Agent Behavior Monitoring (ABM) layer: a four-part infra that captures permissioned trajectories, buckets and analyzes interaction cohorts with metadata, mines implicit user preferences to discover operational rubrics, and converts those rubrics into reliable judge scores and reward signals for training (e.g., RL or fine-tuning pipelines). Practically this requires LLM-driven analysis at scale, custom judge logic, continuous sampling to track distribution drift, and integration with training frameworks (Fireworks, OpenAI, Tinker). When done right, ABM turns production data into interpretable metrics, real-time alerts, and reward models — aligning product improvements with what users actually value and creating a compounding product advantage.