Astra 长上下文下代理五阶段架构拆解
视频主张 100 万上下文不等于可靠代理，关键是把规划、工具调用、验证、记忆、输出分层隔离。
GPT-6 ASTRA MAKES ONE AGENT MISTAKE IMPOSSIBLE TO IGNORE a 1M+ context window sounds massive, but long context alone does not make an agent reliable. what matters is how planning, action, checking, and memory are separated. plan → tool use → verification → memory → output 5 stages keep one bad decision from silently reaching production, while useful results can be stored for future runs. long-running work starts looking completely different. instead of one giant prompt carrying everything, each layer gets a narrow job and a clear failure point. better agents are starting to look less like smarter chatbots and more like small operating systems built around a model.
这条是引用着另一条写的：x.com/gippp69/status/2097696163424014406