, ,

A close look at WAIC | The first light of the agent age, and the distance from reality

Introduction: From “a model’s flourish” to “getting the thing done”

In the past year, the industry’s narrative has quietly turned from “how strong is the model” to “can an agent get the thing done.” The 2025 World Artificial Intelligence Conference (WAIC) was especially typical: the stage no longer focused only on parameters and leaderboards, but on whether an agent can be embedded in a real process, connected to existing systems, and, under a framework that is observable and governable, produce a continuing commercial return. This drift of narrative forces an enterprise to move its attention from “the purchase of a single-point tool” to “the building of a process middle platform and a system of governance.” In short, AI is no longer only a layer of effect stacked on a screen. It is going into the joints of finance, operations, and legal compliance. It is no longer an improvised performance of a flourish. It has to stand the long test of a production line, an SLA, and an audit.

A look back in 2026: this piece was written at the node when the agent narrative had just turned toward “being operable.” Looking again a year later, what truly opened the gap was not a demo, but whether the intent chain, the measurement panel, and the escalation of exceptions had already become muscle memory.

NO° 01 | The agent boom is rising, but what is truly accepted is “being operable”

On this year’s floor you can clearly feel a consensus: for an agent to land inside an enterprise, the key is not “can it do it,” but “can it keep doing it.” This makes both the supply and the demand sides move their center of gravity from a feature list to an operating capability. The supply side has begun to take “an intent chain, a data contract, a permission ticket, rollback, and audit” as the standard kit of a proposal, because without these there is no talking of long-term running. The demand side has also moved from a one-off license purchase to a combined contract of platform and governance services, and has written metrics such as latency, first-resolution rate, interception rate, coverage of citation, and completeness of write-back into the KPIs, so that “is it good to use” becomes “does it count.”

This shift is rewriting the view of process inside an enterprise. In the past we used a process map to train people to follow an SOP. Now it is more like training a “work model”: the same intent chain can be A/B’d, can have a regression test, and can, because of an anomaly, raise the grade of human intervention. When a process has been made into data and can be replayed, management no longer depends on “who signs off.” It lets evidence automatically channel the decision. And at the level of procurement, a platform-style supply has begun to rise; an enterprise can assemble flexibly between a commercial product and an open-source base, and avoid the risk of a black box and of lock-in. This directly gives birth to three new kinds of role: an intent architect who turns a business problem into an executable intent chain; a governance officer responsible for the consistency of standard and of semantics; and an agent-operations person responsible for logs, versions, permissions, and rollback. When these roles are institutionalized, an agent can possibly “grow annual rings” inside the enterprise.

The deeper effect on the commercial environment is that the logic of decision moves from “launching it in one go” to “continuing operations.” An enterprise will be more inclined to start from two or three cross-department pain points, and use a quarter to verify the feasibility of the division of labor between person and machine, the governance of permission, and a closed loop of write-back; if the effect can be proved, then expand to an adjacent scene. This rhythm of “run fast in a small field, and let the metrics speak” replaces the old risk engineering of a huge investment and a one-off reconstruction, and also lets the senior layer bring an AI investment more naturally into the governance framework of the P&L and of cash flow.

NO° 02 | WPS AI: how a native office agent rewrites “the economics of the document”

In an office scene, the value of a native agent is not in how well it writes, but in how well it connects. When writing, reading, table analysis, and the design of a deck are strung into a conversational workflow, and interlock with the enterprise knowledge base, an image library, data sources, and APIs, the transaction cost of “making a document” will fall sharply. Before a meeting, the system automatically generates a summary for review on the basis of pre-meeting materials; during the meeting, it archives passage by passage and annotates resolutions; after the meeting, it tracks action items and owners, and writes citations back to the knowledge base. From a management point of view, weekly and monthly reports are restored as “a carrier of evidence and insight,” not a physical project of layout and typography. From a finance and people point of view, low-value clerical work is taken on by the model, human resources turn toward decision and collaboration, and time is released to tasks that truly have incremental value.

This kind of product has three deeper effects on an enterprise. First, knowledge is no longer a static document. It is “servitized” in the manner of semantic retrieval. The period for a new person to get up to speed is shortened; cross-department Q&A is completed automatically through citation and provenance, reducing the hidden cost of “going to ask someone.” Second, compliance work is forced to the front: brand tone, sensitive words, and rules for sending outward are institutionalized, and provided to the generation module in the form of rules, bringing the probability of “stepping on a mine” down to an acceptable range. Finally, the culture of management is quietly calibrated: meetings and reports no longer compete on beauty, but on citation, on consistency of standard, and on the effectiveness of insight. This also brings the evaluation of performance back from “the number of pieces and the number of pages” to “the quality and the speed of a decision.”

Of course, a real boundary is still there. Long-context reasoning, consistency across files, and the handling of brand detail still, in the short term, need a second human review; privatization and a recasting of permissions need a budget in advance and cross-department collaboration; pushing everyone to learn conversational office work also has a learning curve. A feasible method is first to build confidence with low-risk, high-frequency tasks, such as meeting minutes, a first draft of a report, a first version of a deck, and at the same time build two gates of “generation must cite” and “high risk must have a second review.” When the metrics (document-completion latency, first-pass rate, coverage of citation) improve month by month, the enterprise will naturally hand more processes over to the agent to continue and complete.

NO° 03 | Baidu’s Huiboxing NOVA: the supply curve of a digital-human livestream, and a paradox of trust

A digital-human livestream is not as simple as replacing a person with a virtual image. It is a reshaping of the supply curve of selling. The supply of content for long hours, nighttime, and long-tail SKUs used to be constrained by a live host’s stamina and calendar; now, at the same cost, you can obtain a more stable coverage and rhythm. The logic behind it is script generation driven by a large model, real-time decision on the rhythm of interaction in the livestream room by a multi-agent system, and a highly realistic sound and picture that keeps a persona consistent. For retail and brands, this means that “content” moves from a scarce asset to a craft that can be produced at scale, and the importance of an operations middle platform rises with it — what can truly be copied is a product knowledge base, a script FAQ, a strategy of interaction, and an SOP for handling a crisis, not the persona itself.

But a digital human also faces a paradox of trust. The reason leading cases succeed is mostly that they carry a strong persona and a social trust that already existed, which significantly lowers the psychological cost of “endorsing” and of “selling.” An ordinary brand, if it hopes to replace “persona assets” directly with a digital human, will often overestimate the short-term effect. A more realistic strategy is first to position the digital human as “a fill-in for a time slot and a category,” let it take on long-tail and low-margin SKUs, do a settling into knowledge and A/B tests in the operations middle platform, gradually optimize the script, the expression, and the rhythm of offers, and then, selectively, grow a persona double for the brand. This path can let content capacity and the settling of trust accumulate in sync, and avoid a homogenization of content and an aesthetic fatigue caused by “seeking a burst in the short term.”

Compliance is another clear boundary. The subject of responsibility for a virtual endorsement, advertising disclosure, the purpose of data, and the rights and authorization of likeness and of voiceprint must be clearly agreed with the platform and the vendor before going live, and an audit and a rollback mechanism must be embedded in the process of every broadcast. In experience, a good digital-human livestream is like a factory: the data layer and the strategy layer at the front must be solid; the interaction and the price rhythm in the middle must be monitorable and hot-updatable; the write-back and the summary at the back must feed the next broadcast. When this closed loop can run steadily, a digital human can possibly evolve from “a fresh gimmick” into “a capacity that can be expanded.”

NO° 04 | JD Cloud JoyAgent: an open-source multi-agent system rearranges the calculus of “build vs. buy”

An open-source base for multi-agent systems is turning an enterprise’s choice of technology from a binary into a question of assembly. When an enterprise-grade agent framework is provided in an open-source way, including end-to-end task orchestration, the mounting of sub-agents, and modules of observation and of permission, an enterprise can, at a lower cost of trial and error, first get a key process running, and then, according to maturity and a preference for risk, decide where to buy commercial, where to build, and where to mix. This changes the structure of bargaining power: the buyer is no longer passively accepting a black-box product, but can use “an open-source base + private capability” to form a dual backup of technology and of supply, and also forces the vendor to turn the value proposition from “a magical effect” to “a service of governance and operations.”

This kind of technical supply also recasts the division of labor in the ecology. Systems integrators and consultancies can, around a general base, provide an industry packaging and a recipe of governance, rather than reinventing the wheel. A platform team and a DevRel role inside the enterprise need to understand a multi-agent framework and a semantic interface, and thereby connect people, data, and tools into a net of capability that can be operated. What is worth watching is that open source is not equal to being free of operations. When an enterprise hands more key tasks to a self-built agent, AgentOps capabilities such as observation, permission, version, disaster recovery, and the escalation of exceptions are a necessary investment. A truly stable plan is often not “the most features,” but “when it goes wrong, you know where it went wrong, and where you can go back to.”

At the layer of connecting to old systems, the most pragmatic method is not a large integration, but first laying a “semantic bridge.” Vectorize documents, images, and tables into a retrievable corpus, use a permission ticket and a provenance mechanism to control calls and citations, and then gradually expand to real-time reading and writing. This can, without tearing down and rebuilding, let an agent move among old systems, both lowering the risk at the start and keeping the elasticity of iteration.

NO° 05 | AICE ADGo: recast “workshop creativity” into “a production line + a hand finish”

The difficulty of marketing creativity is not in producing an idea, but in producing enough good ideas that can be executed, can be evaluated, and are consistent in style. After multi-agent collaboration cuts “research—insight—strategy—script—visuals” into parallel modules, the capacity of the first round can be expanded very markedly: the same brief can, in a short time, receive several strategic drafts, arguments, and visual samples, and a meeting of the brand and the agency moves from “out of stock” to “picking from stock.” This way of supplying changes the economics of the creative industry, and lets small and mid-size enterprises, for the first time, obtain a whole case at a “near-4A” level at a cost they can bear; the time of decision is also compressed because of “a comparison of many versions.”

The true difference, however, still needs a person to define it. AI is good at integrating data and summarizing a pattern, but finds it harder, on its own, to step over inertia. Brand tone, cultural detail, humor against common sense, and a turn of style must have their boundary drawn by a person. The truly effective way of working is therefore not to discuss “which idea is good,” but to discuss “on which dimensions we want to deviate from the average.” Under this structure, AI’s role is to pull capacity to the full, prepare the evidence, and spread the versions; a human’s role is to make those trade-offs that carry a style. When the mechanism of media spend and of write-back is sound, the market performance of the winning version will feed back into the knowledge base, and the next round of generation will carry a real memory of the market, and not rely only on the model’s corpus intuition.

The governance of data and of copyright is the premise of whether this kind of platform can enter a core project. Brand materials, historical sales data, and user corpus must be used under clear terms, and, when necessary, deployed privately or in isolation. At the same time, all output should come with citation and version information, to ensure provenance and accountability. When data and governance are in place, a creative production line of human–machine collaboration can steadily deliver a feasible plan to the market end.

NO° 06 | SenseTime’s “Office Raccoon”: let a non-technical post have “the right to self-serve analysis”

For most large organizations, data and knowledge are scattered in countless folders, reports, and systems. What is most expensive is not storage, but obtaining it. A knowledge agent that, in a conversational way, takes on “task planning, data upload, cleaning, analysis, visualization, and Q&A” is freeing non-technical posts from the passivity of “waiting for a report, looking for a colleague.” Planning, purchasing, and operations can, in natural language, ask “which category’s gross margin is changing” and “which region’s returns rate is abnormal,” and obtain a visualized answer by themselves. The role of the data team therefore rises to governance and insight, and is no longer a factory of repeatedly producing charts.

This turn toward a tool will produce several marked effects inside the organization. First, the knowledge junkyard is emptied: once SOPs, institutions, and historical close-out reports are vectorized, they can be recalled by semantic retrieval, and attached to the output in the form of a citation, forming a positive cycle of “settle while using.” Second, the rhythm of decision is pulled faster, but a buffer also needs more to be designed. What AI gives is an insight of “look first”; a person still has to bear responsibility for causal relations and for the consequences of a strategy. In a high-risk decision scene, therefore, a dual-track process (an AI suggestion and an expert review in parallel) is a necessary safety valve. Finally, the governance of security and of privacy has to be completed in advance, including an inventory of data, rules of desensitization, role-based permission control, and an audit of queries, or “good to use” will quickly part ways with “usable.”

From the point of view of commercial value, pushing down the right to self-serve analysis will raise the overall density and speed of decision, let more micro-decisions be completed on the spot at the front line, and reduce the traffic cost of the middle layer; at the same time, it will also raise the requirement of data literacy. When more people can ask the data, the enterprise must teach everyone to read the data, to doubt the data, and to correct the data, or “blind faith in AI” will lead to a new bias of decision.

NO° 07 | A common law across modules: operable, observable, governable

Spread out the plans from different exhibition areas, and they finally converge into three common laws. The first is being operable: an enterprise has to treat work as an intent chain, not a static process, and, for every intent chain, define the goal, the data contract, the human–machine boundary, and the rule of write-back, so that it can accumulate the ability of “having learned to do the thing.” The second is being observable: without measurement, everything is a feeling. The reliability of citation, the latency from intent to result, the share of human interception, first-resolution rate, and the share of steps completed automatically — these metrics bring a management discussion back to the facts. The third is being governable: a grading of permissions, an audit that can be traced, and a version rollback are the underlying engineering that upgrades “can run” into “can run with ease.” Once something goes wrong, if the system knows how to downgrade and roll back, the team will have the courage to hand more tasks to the machine.

Closing: The first light is already here, but what truly shines into the P&L is the power of governance

The bustle of WAIC is only for a moment; the operations of an enterprise are for a lifetime. Whether it is a native office agent, a digital-human livestream, an open-source multi-agent base, a recasting of a creative production line, or a self-serve analysis from knowledge to decision, they are all, in essence, testing the same thing: whether we have the governance capability to turn AI into “a second work queue.” When an intent can land, data can be cited, a chain can be observed, and risk can be governed, AI can move from a demo on the stage to a curve on the financial report. Rather than chase the next splendid feature, first run one intent chain steadily, put up one measurement panel, and polish one process for the escalation of exceptions. When these three things become muscle memory, the first light of the agent age will naturally shine into the enterprise’s P&L.

More deep-dive pieces

The judgment, restated

My judgment is that the first light of the agent age is already here; what can shine into the P&L is being operable, observable, and governable — not the next flourish.

Three things this week

  • Pick one cross-department intent chain, and write clearly the goal, the data contract, and the human–machine boundary.
  • Put up a minimum measurement panel (latency / first resolution / interception / completeness of write-back).
  • Set one path for the escalation of exceptions: when a person must sign off, and when to roll back.

You are welcome to subscribe to Future Insights Lab, or to reply and tell us: which intent chain will you first run steadily this week?

訂閱 Future Insights Lab

不定期寄送 AI × Business 短篇筆記,協助你掌握實務重點。隨時可取消訂閱。