Why RL-Trained Agents Fail in the Real World — and How to

RL-trained coding agents ace benchmarks but stumble on messy real-world interfaces. Demos show agents guessing passwords, clicking ads, and filling forms on wrong sites. Partial observability, irreversibility, non-determinism, and adversarial content break the tidy RL assumptions.

The fix is not better models alone — it's training in high-fidelity digital sandboxes that simulate layout shifts, slow loads, pop-ups, and expired sessions. Process reward models penalize dangerous paths, not just outcomes. Recovery must be a native action, and adversarial tasks should be mainstream training.

The model needs grounded perception, change detection, and multi-source observation. The harness — the interface between model and world — adds checkpointing, risk classifiers, credential guardrails, execution monitoring, and forced human handoff when confidence is low.

The bottom line: the difference between a demo and a product is what happens after the first failed click. Simulate reality in training, let agents fail, and keep the harness strong early. Over time, the model improves and the harness thins.