InfoQ reports on Xiaohongshu's enterprise AI assistant Seal, based on a presentation by company architect Meng Hongjin. The team started with three people and delivered an internal beta in three days using OpenClaw as an open-source base rather than building an agent platform from scratch. After an initial 200-user test, the team upgraded the architecture for horizontal scaling, opened 2,000 public-beta places, and reported that all places were claimed within five minutes. The article says company-wide access was achieved in mid-March, roughly one month after the initial effort.
Seal runs entirely in the cloud rather than on employees' personal computers. InfoQ describes three security layers. At the cluster layer, agent workloads run in Kubernetes clusters isolated from production networks while retaining controlled access to the office network. At runtime, the initial design used one Pod and a dedicated cloud disk per user to preserve each OpenClaw workspace and state. Nex adds lifecycle orchestration, warm pools, and stronger isolation. Its reported stack includes an AI gateway for identity and traffic control, a control plane for sandbox lifecycle, custom Kubernetes CRDs and an Operator for layered scheduling and batch creation, and a microVM security layer built on Kata Containers and Cloud Hypervisor. Controls include system-call allowlists, network and filesystem restrictions, cgroup resource limits, and security auditing. Because internal GPU capacity is limited, Seal also uses external cloud model APIs alongside internally hosted models. For data protection, a security SDK built into the Seal client scans uploaded files; content identified as sensitive is restricted to internally hosted models rather than external model APIs.
The team describes two cost controls. Self-GC treats long agent histories, including tool-call results, as a structured context heap rather than trimming text indiscriminately. When context reaches a threshold, it asynchronously proposes a cleanup plan, previews the expected token savings and boundary risks, then waits for a safer execution point such as completion of a tool call or a pause for user input. InfoQ reports that after launch, daily token consumption fell by 10–15%, with peak reduction of about 20%. These figures are reported by the team, not independent measurements.
Seal's Auto mode selects among model tiers. The article says a 30B mixture-of-experts model estimates task difficulty, with rules that also account for cache hits and fallback behavior if a router or recommended model is unavailable. In online canary data, the team reported that over 70% of daily tasks could use lower-cost models, about 22% needed a higher tier, and about 3% needed a top-tier model. Compared with routing all users to a higher-tier model, the team reported approximately 69% lower model costs.
Memory was prioritised over self-improving Skills after users reported that the assistant forgot earlier conversations. The team found OpenClaw's native memory depended on users explicitly asking it to remember, lost context after compaction or a new session, and promoted little to long-term memory. Seal's memory has three layers. The session layer loads SOUL.md, key profile information and preferences into the system prompt. When a conversation is compacted, a new session starts, or a nightly scan runs, facts are written to a worklog that serves as a traceable evidence buffer and is not passed to the agent wholesale. A third layer continuously compiles the worklog into a structured, linked Memory Wiki rather than relying on conventional RAG: profile-critical entries are always injected, while the rest is retrieved on demand, with the model given the wiki's table of contents and using file-search tools to look up entries. The memory still lives in local files; the stated next steps are cloud-based memory shared across agents and project-, repository- and organisation-level shared memory.
The article also reports an internal Skillhub ecosystem with more than 9,000 Skills after slightly over a month, including over 8,000 internal Skills. Skillhub distinguishes platform Skills maintained by owning teams (about 100), team Skills (about 136), restricted private-domain Skills for sensitive systems such as HR and finance, and personal Skills, which are the majority; most calls go to the top 100–200 Skills. A companion service, CoWork Studio, lets employees publish tools they build with AI coding as shared internal websites. These are platform adoption figures and are not presented as deployment outcomes.