Alibaba Cloud's Agent Technical Expert Zheng Tian published an analysis on October 27, part of the open-source Alibaba Cloud AI Agent Handbook, of the personal agents released since August: Meta Muse (September 8), Manus 2.0 with Cue (September 28), OpenAI dots (September 29) and Grok Bot (from August 11). The article divides the stack into four parts: a runtime kernel pairing a model with a harness, a dedicated execution environment such as its own Linux desktop, a tool and service ecosystem, and a digital identity made of accounts, payment instruments and phone numbers. Its central claim is that an always-on agent is a runtime change. Task ledger, memory, permissions, credentials, triggers and completed external actions live outside the conversation, so a new email or calendar event can wake the agent on its own machine. The author separates persistent state from retrieval, citing FlowState (arXiv:2609.34565), which stores typed state nodes, their relationships and references to raw tool output, and reports about 4.6 and 14 percentage points higher success on MemoryArena and t3-Bench with roughly 40% fewer tokens.
Our takeThe useful distinction here is between remembering more and keeping the right objects: state nodes, their links, and pointers to raw tool output, rather than a longer transcript. The reported gains come from one paper's own benchmark run, so they are a design signal, not a settled result. The harder problem the article names - memory that must consolidate, compress and forget, with payments and messages that stay auditable - has no shared evaluation yet. The thing to watch is whether these products expose state provenance, not whether their context windows grow.