Long agent sessions never slowed down because the model got tired. They slowed… · @harden_369741 · TwKit

@harden_369741

Long agent sessions never slowed down because the model got tired. They slowed down because I made the model retell the whole history every single turn. Anthropic just put message threads in public beta on the Claude Platform. An agent sends only the new messages, and median TTFT falls about 35% on 500k to 1M token requests. Claude Code turns it on by default for most users, and it works on native Claude Platform plus Anthropic managed versions on AWS, Microsoft Foundry and Google Cloud. My rule is plain. Put long sessions on threads, and keep the session alive inside 5 minutes, or the thread expires and you resend the full history anyway. Treat the 35% as a latency win, not a saving, because standard token rates and rate limits do not change. A good long session is one that stops repeating itself. https://claude.com/resources/articles/message-threads #Claude #AIAgents #ClaudeCode

曝光 26 · 评论 0 · 点赞 0 · 书签 0