# Decagon explains RL for a production customer-facing agent at hundreds of billions of parameters

> Original title: 3 | @DecagonAI

- Source: LMSYS Org
- Published: 2026-11-05T22:43:26.000Z
- Bad Influence: https://badinfluence.news/items/vp4p73trygids7kmjk1ezm4k1
- Original: https://x.com/lmsysorg/status/2118474298541457519

## Summary

Decagon's Cyrus Asgari discusses synthetic data, reward design, and stabilizing reinforcement learning under training-inference mismatch for customer-facing agents. The post presents these methods as ways to build better production agents, with no model release or benchmark results given.

## Post

3 | @DecagonAI 
RL for a Production Agent at Hundreds of Billions of Parameters

Cyrus Asgari on synthetic data, reward design, and stabilizing RL under training-inference mismatch to build better customer-facing agents.

![](https://badinfluence.news/api/img-proxy?u=https%3A%2F%2Fpbs.twimg.com%2Fmedia%2FHWZTALMaYAAVave.jpg&mode=thumb&exp=1794182400&sig=91be9ea6aeb4e367)