@VictorTaelin The Codex (or Claude Code) harness does context compaction, and they're surely training for this on two levels: the model is being trained within harnesses that are controlling their context, and the models they're using to compact context are also being RL'd to perform better.

Source

• URL: https://x.com/theunabletable/status/2078150286233760155
• Author: drew (@theunabletable)
• Posted: 2026-07-17 16:10:13

Branch

1/ @theunabletable

@VictorTaelin

The Codex (or Claude Code) harness does context compaction, and they’re surely training for this on two levels: the model is being trained within harnesses that are controlling their context, and the models they’re using to compact context are also being RL’d to perform better.

Related

• Spine: 2026-07-17-read-this-and-your-next-model-will-be-10x-smarter