x-octo home Business judgment on AI products
中文

VOL.2026.10.16 Today's call 4 min read

Personal errand agents and agent security both heat up: assistants start making calls and placing orders, while engineering teams start locking agents down.

Friday, October 16, 2026

—
Personal errand handling becomes a new battleground: Instinct takes requests by text or call and uses a phone and computer like a person to order groceries, book tickets and send flowers.
—
Meta Muse pushes a general agent to the consumer side, but the public material offers only one positioning line; its flow and deliverables remain unverified.
—
Agents now actually touch files and run commands, so agent-security and agentplay split off tool-call policy and session replay as their own steps.
—
Hong Kong customer service complaints surpassed the full-year 2025 total, partly attributed to enterprise use of AI customer service, as delivery quality and acceptance start to bite into adoption.
01

Today's positive direction

The clearest line today: assistants are moving from giving advice to executing on your behalf. Instinct hands execution to the model to make calls, place orders and book tickets; Meta Muse pushes a general agent to the consumer side; meanwhile tools like agent-security and agentplay handle what happens after an agent actually touches files and runs commands. Once execution is handed over, boundaries, auditing and replay become necessities — that is the direction worth watching today.

02

Opportunity roundup

3 picks
01

Personal errands: from advice to execution

Instinct's shape is concrete: users reach it by text or call, and it uses a phone and computer the way a person would to order ingredients, book a weekend train home, buy groceries, send flowers to a friend or find a weekend venue. What the user gets is an executed errand result, not a to-do list. The value is handing execution to the model, entering through high-frequency, checkable everyday tasks.

Meta Muse takes another path along the same trend: a consumer-facing general agent that claims to get things done for the user. But the candidate material offers only this positioning, with no detail on who opens it at which work step, what material the AI receives, which actions it performs or what it finally delivers — flow and deliverables remain unverified.

02

After agents start moving: security and replay become their own steps

agent-security targets engineering teams: before letting coding agents call tools, they use it to set tool-call policies on-device, detect leaked secrets and log agent actions, ending up with an auditable local record. The exact policy configuration flow and interception effect remain unverified.

agentplay handles the aftermath: after a coding agent finishes a task, developers feed the session record into it, which replays it as a map and timeline so they can locate the step where the agent went wrong. Which agents are covered and how readable the replay is remain unverified.

Together they show one thing: as coding agents take on multi-step tasks, the tool-call layer and the process-review layer are each being carved out as separate work.

03

Vertical workflows carved out on their own

3DDX claims to use AI for dental implant planning, but the public material is only a launch headline, with no detail on which images the AI receives or what plan files it outputs. Implant planning is one of the few dental steps that is both image-reading heavy and structurally decomposable, suggesting vertical imaging workflows are being carved out separately.

Qianwen Office E-commerce Edition claims to hand store-opening steps — product selection, store setup, listing and customer-service scripts — to AI, opening a store in seven days for a few hundred yuan; the report does not say what the AI receives or what materials it produces. The big cost of opening a store has long been labour and trial-and-error rather than software subscriptions, suggesting e-commerce services are shifting from selling tools to selling the result of getting a store open.

Actually Asked targets brands: it turns common Q&A into a branded public answer page, with the AI taking that Q&A content and producing a page that answers visitor questions on its own. The update mechanism and delivery form remain unverified.

adu-motion-video targets talking-head or explainer creators: they download a .motion template from MotionGit and hand it straight to their AI assistant, which generates code-level animated segments. How templates and assistants work together remains unverified.

03

Market context

Hong Kong customer service complaints surpassed the full-year 2025 total, with the report partly attributing this to enterprise use of AI customer service. This reflects delivery-quality and acceptance problems for AI customer service in real service scenarios, and may affect enterprise adoption pace and compliance requirements.

On the compute side, CoreWeave's bond yield exceeded 15% and its shares fell nearly 15% intraday, with an analyst warning its high-debt growth model is unsustainable. The debt-driven GPU compute expansion model now faces financing-cost and credit concerns, which may raise funding costs for AI compute suppliers.

On the platform side, Apple is bringing its first AI features to the China market — a platform policy and ecosystem change, not a standalone product; the public material offers only the timing that it is planned for December after a delay, with feature scope, partners and developer interface terms undisclosed.

04

Conclusion

Today's opportunities cluster in two places: handing execution to the model for personal errands, where Instinct and Meta Muse represent concrete errands and a general agent respectively; and the supporting layer for agents that actually move, where agent-security governs tool-call boundaries and agentplay governs process replay. Vertical workflows being carved out (3DDX, Qianwen Office E-commerce Edition, Actually Asked, adu-motion-video) is the same logic extended to industry. The market context is a reminder: AI customer service delivery quality is already biting into adoption pace, and compute-side financing costs are rising — both will shape who executes solidly next.