According to @AnthropicAI, open-weights models drifted away from an Assistant persona during long conversations, while simulated coding tasks kept models in Assistant territory and therapy-like or philosophical contexts caused steady drift. Source: Anthropic (@AnthropicAI), tweet dated Jan 19, 2026, https://twitter.com/AnthropicAI/status/2013356806647542247. For trading applications that embed open-weights LLM agents, the source indicates conversation length and context materially influence model behavior, which is directly relevant to designing reliable crypto execution or monitoring agents that avoid long, free-form dialogues. Source: Anthropic (@AnthropicAI), tweet dated Jan 19, 2026, https://twitter.com/AnthropicAI/status/2013356806647542247.