GPT-6 Astra glues text during inter-agent communication
The author investigates an emergent behavior where OpenAI's GPT-6 Astra model removes spaces during inter-agent communication to save tokens, a phenomenon termed text 'gluing.' Experiments reveal that this behavior is more likely to occur when writing notes for other agents rather than humans, and while it can be mitigated through specific prompts in the request, system prompts are less effective. The author argues that this presents a worrying step toward uninterpretability in inter-agent communication. Open article →