Probability, Power Outlets, and 5,000 Einsteins: An Interview with Alibaba Cloud R&D Vice President Zhou Qi

Probability, Sockets, and 5,000 Einsteins: Interview with Alibaba Cloud R&D Vice President Zhou Qi

Spending money but not gaining efficiency? Security and organization lagging behind? Systems not stable enough, and when the model changes, the entire architecture has to be rebuilt?

Alibaba Cloud Intelligence Group Vice President of R&D and head of the Cloud‑Native Application Platform, Zhou Qi, didn’t pick just one. He talked about the development of electricity.

“In past industrial revolutions, power generation relied on large power plants and a single massive transmission shaft, with energy output being centralized. You can’t have every household building a huge mechanical generator at home.”

Electricity makes energy usage egalitarian. AI makes intelligence egalitarian.

The drive shaft is a concrete object in industrial history. In American factories at the end of the 19th century, a steam engine drove a long shaft that ran through the plant, with belts hanging from the shaft to connect machine after machine. After electric motors entered factories, many companies initially simply replaced the steam engine with a large motor while keeping the shaft and belts unchanged.

Economic historian Paul David noted in his 1990 study that by 1899, electric motors accounted for less than 5% of mechanical power in U.S. factories; it was not until the 1920s that electrification began to show a significant impact on manufacturing productivity.

Today’s enterprises also have their own “drive shafts.”

The code, programs, and systems accumulated over the past three to four decades, along with team collaboration methods and management philosophies, are largely built on formal logic. Traditional software is a symbolic system where inputs, conditions, and outcomes should, in principle, be deterministic; but AI operates on a probabilistic foundation.

“There is an inherent paradox here. That is the essence of the matter.” Zhou Qi said.

An individual uses AI to generate an image, sees a slight pixel shift, laughs it off and moves on. A company invests funds in a scenario, only to wait for returns that may come a year or even five years later. It must answer: can the system run continuously, and can it ultimately generate a return?

“A business is an ROI scenario; it requires certainty.”

Zhou Qi joined Alibaba Cloud as the first intern of the Feitian team, participating in the development of the Feitian operating system from zero to one. Since then, his work has gradually expanded to log processing, system monitoring, and diagnostics.

This interview took place after the 2026 Yunxi Conference. During the conference, Alibaba Cloud launched the intelligent agent building and governance platform AgentCore, further standardizing and hosting enterprise‑grade Agent infrastructure.

The concerns mentioned at the beginning—efficiency, security, organization, and stability—actually all point to this dilemma. How to establish certainty over probability has become the issue that Agents must solve to enter production environments.

Extract power from the socket

When electricity first entered American households, wiring was primarily installed for lighting. If people wanted to use an iron, fan, or vacuum cleaner, they often had to unscrew the light bulb first and then plug the appliance into the lamp socket.

In 1904, Harvey Hubbell received a patent for a “detachable plug.” The plug could be pulled out directly, but the matching base still had to be screwed into the lamp holder.

The real issue quickly shifted from “whether there is electricity” to “how to connect to it.”

As wall outlets multiplied, different manufacturers adopted various specifications. Litigation records from 1919 involving Hubbell, General Electric, and others show that around 1915, the U.S. market had 15 to 20 types of plug blades and 15 to 30 types of sockets, making products from different manufacturers incompatible. The ruling also clearly documented the urgent market demand for standardization at the time.

Zhou Qi mentioned electricity for the second time, referring precisely to this set of supporting facilities.

“For the first 30 years, everyone used a single lighting system. It wasn’t until the specifications for electricity, as well as standardized relays, transformers, sockets, and other accessories, were established that a very prosperous economic cycle emerged.”

Agents within enterprises have reached a similar stage.

If there is no common foundation, each team can only develop models, write tools, and configure permissions within its own “small workshop”: today building an “electric boiler,” tomorrow recreating an “electric fan.” The identities of different agents cannot be recognized, permissions cannot be inherited, and systems cannot connect smoothly. In the end, many demos are produced, but they struggle to enter real production environments.

This is not a problem unique to any single vendor. As agents move from experimentation to production, industry focus has shifted from “whether we can build agents” to how long-running tasks can operate continuously, whether state can be restored, how credentials are managed, and how behavior is audited. Accordingly, the runtime and governance infrastructure that hosts agents is becoming a common enterprise requirement.

Enterprises urgently need this infrastructure because they are delegating more work to AI.

Previous AI applications were mostly focused on isolated efficiency gains: developers used them to assist coding, product managers used them to generate PRDs, and marketers used them to organize information. As models become better at handling complex tasks and invocation costs drop, AI is moving from peripheral roles into the core production processes of enterprises.

“When agents become stronger than humans and cheaper, everyone will shift AI applications from marginal scenarios to the most critical production scenarios in enterprises.”

But once an Agent enters the core workflow, the ad‑hoc, workshop‑style approach can no longer sustain it.

“This isn’t a lack of industry insight on the part of enterprises. On the contrary, companies often know their own business better than the platforms do,” says Zhou Qi. After AI capabilities become widespread, the truly scarce skill will be another one: planning system architecture for long‑term evolution after having seen enough industries and scenarios.

Every functional department knows how to improve the work it handles. The harder part is making those capabilities connect—so that a task completed by one department can be taken over by another; tools and data used by one Agent can be reused by another Agent within permission limits; local efficiency gains can ultimately be integrated into a complete production workflow.

This architecture needs to address four issues: stability, security, performance, and cost.

Stability is not just “no service downtime”; it also means that AI operates along the paths required by the enterprise while exercising its capabilities. Each enterprise represents an abstraction of its specific industry, and the system must translate this abstraction into rules, boundaries, and evaluation criteria.

Security must protect the foundations on which the enterprise has historically survived: job structures, permissions, compliance, and privacy. Traditional software controls deterministic program calls, whereas an Agent may autonomously plan and continuously invoke multiple tools. Once it enters the production workflow, the scope of security governance expands from model outputs to identities, permissions, data, tools, memory, behavior, and runtime environments.

In the electricity analogy, Alibaba Cloud’s open‑source AI Infra and related components are like existing socket standards. Enterprises can use them directly to develop their own “refrigerators” and “electric fans.” AgentCore is a one‑stop, enterprise‑grade platform for building and governing Agents. Companies don’t have to start from wiring, fixing transformers, or implementing power‑off protection—they can focus their energy on specific business needs.

Standardizing infrastructure does not mean that a company’s existing software systems will be discarded.

On the contrary, the processes, interfaces, and data accumulated over time in legacy systems are precisely the “hands and feet” that AI needs to be effective.

“If AI doesn’t have hands and feet, it can’t fully deliver value,” says Zhou Qi.

One Person Monitoring Five Systems

The Manbang Group case can explain what these “hands and feet” actually are.

Manbang’s platform connects shippers and truck drivers. In 2025, its platform fulfilled 236 million orders, a year‑over‑year increase of 19.8%. For a platform that has become a fundamental part of the logistics infrastructure, a single system failure can mean vehicles cannot be dispatched and cargo is delayed.

A dispatch must pass through the truck side, the driver’s phone, the shipper’s app, and the backend dispatch system. Any hiccup in the network, database, or application service can cause the task to fail.

People troubleshooting the issue have to follow this chain and search layer by layer for the cause. According to Manbang’s case material, an operations engineer monitors five systems and must open six platforms during an emergency. Diagnosing a fault requires switching between different platforms, querying them, and then piecing the scattered information together.

“Fixing sometimes only takes five minutes, but finding the issue takes half an hour,” said an engineer who has done stability on‑call for years on similar platforms.

AgentCore takes over exactly this part of the work.

The first step is to turn the company’s existing operational troubleshooting tools into MCPs, converting them into standard interfaces that the Agent can call. The second step is to bring the existing permission system along as well.

Taking a machine offline, restarting an application, modifying a configuration— the enterprise already has corresponding roles and approval processes. Who has permissions, what level of incidents can be handled automatically, which actions must be confirmed by a person—these rules do not disappear just because the executor changes from a human to an Agent.

After completing these two steps, the operations system is no longer driven entirely by humans; it now has a “brain” that can run around the clock, complemented by the hands and feet for querying, diagnosing, and handling faults, capable of executing thousands of commands within a minute.

According to Manbang’s presentation at the Yunxi Conference, its stability digital human can pinpoint the cause of major risk scenarios within one minute, with an accuracy exceeding 90%.

“In the past, operations staff would receive calls in the middle of the night, log into the system, and mainly perform emergency actions. This type of work doesn’t necessarily foster growth, yet it constantly consumes people’s attention,” said Zhou Qi. Now they can focus on more important tasks.

The source of the tools and the extent of task completion still need to be defined by the enterprise.

Zhou Qi refers to business processes and data collectively as the “enterprise context.” Each industry’s context is different, and this part cannot be replaced by a generic platform. Enterprises need to accomplish four things: define tasks, define tools, define the Agent’s identity and orchestration method, and prepare the context required for its operation.

The platform is responsible for the common issues.

Some production tasks need to run continuously for days and nights and cannot be interrupted by a single machine failure; after a task restarts, it must not lose its previous state. Which data can be accessed, which operations can be performed, and where to recover from after a failure are all managed uniformly by the underlying system.

Different scenarios also require different construction approaches.

For “highly controllable” tasks, a high‑code approach can be used, writing the critical steps as deterministic logic. “In a sense, this brings us back to the symbolic systems of traditional software,” says Zhou Qi. For scenarios that need to be deployed quickly, Harness can be used, with Skills, rules, and an evaluation framework to constrain the Agent’s behavior.

A bigger shift occurs in task planning.

Traditional software first defines a workflow: what to do in step one, which API to call in step two, and which branch to take when an exception occurs. The AI‑native planning approach is more like cooking a dish—first gather the ingredients, clarify dependencies and the final goal, then set the boundaries of the intermediate process, leaving the concrete steps to be dynamically executed by the Agent.

A year ago, enterprises had limited confidence in long tasks and preferred to use AI to call small tools and solve minor problems. Today, as models improve their support for complex tasks, companies are beginning to define only the goal, resources, and constraints, letting the Agent find the path on its own.

“The latter is more uncertain, but the ceiling is also higher,” said Zhou Qi.

Five Thousand Einsteins

Just how high can this ceiling be?

“A complex task that would normally be handled by a single employee—can it instantly become something that 5,000 employees tackle simultaneously in an emergency? And what would happen if all those employees possessed top‑tier reasoning and execution abilities?” Zhou Qi asked.

“We could even have 5,000 Einsteins orchestrating it. Some do R&D, some do research, some do promotion, and finally we stitch the solution together.”

But the follow‑up question is: when 5,000 Einsteins work at the same time, whose voice is heard? Who can view which data? Who makes decisions when disagreements arise?

A regional operations manager at a chain retail company used to dread inventory reports that came out after 10 p.m. When a store was found out of stock, he had to ask each store manager individually in the group chat and then call them one by one to confirm.

“Sometimes you call and no one answers, and even if they do, there may not be any stock.”

The retail industry has long borne high coordination costs. Stores are scattered across different locations, and each store needs someone to handle replenishment, reconciliation, and operations. Now, this company is trying to build Agent teams for different stores on AgentCore: some roles handle replenishment, some handle reconciliation, and others manage and coordinate.

“This is already more than just an extension of SaaS,” Zhou Qi said.

SaaS mainly brings the physical world into a digital system, making business visible, manageable, and controllable. Agents go further by addressing productivity issues—the system no longer merely records a task, it can also participate in and complete the work.

After electric motors entered factories, the reason productivity only improved noticeably after several decades is also related to this. Initially, factories only used electric motors to replace steam engines, then progressed to one motor driving a set of machines, and finally arrived at the “one machine paired with one motor” single‑machine drive model. Machines were thus freed from the constraints of drive shafts and could be rearranged according to process steps and material flow, and factories gradually shifted from multi‑level structures to more flexible single‑level layouts. The real productivity boost comes from redesigning the production system around the new technology.

In Zhou Qi’s view, the Agent team also needs to carry out a re‑layout, except the subjects have shifted from machines and factories to positions and organization.

In human organizations, if the parallelism of the R&D process is insufficient, companies usually hire more staff; if a particular role becomes a bottleneck, it may be split into Role A and Role B, allowing two people to collaborate to complete the work.

But the splitting and recombination of Agents will be even more frequent. They don’t need to go through traditional handovers and can process tasks concurrently at a faster pace. “Enterprises will start with localized innovation, achieve single-point efficiency gains, and then gradually develop a production model that is AI‑centric with fixed processes as a supplement,” says Zhou Qi.

The specific way they collaborate depends on the task.

Software development needs to link planning, development, testing, and operations, with each Agent representing a role and perspective. For innovative tasks, multiple Agents can brainstorm together with humans, then converge the differing opinions into relatively definitive conclusions.

“It’s a bit like probability. If you lock one person in a house to think all day, they may stay forever in their own mind. But when people with different opinions discuss together, ideas can erupt.”

In Zhou Qi’s view, this multi‑Agent approach will become the mainstream in the future.

However, the more Agents there are, the more complex the management becomes. As the solution space expands, without proper division of labor, evaluation, and exit mechanisms, they may end up duplicating work and consuming each other, ultimately failing to achieve better results.

Security risks also increase with longer collaboration chains. A corporate survey released by the Cloud Security Alliance showed that 65% of respondents had encountered at least one AI Agent‑related incident in the past year, with common impacts including data exposure and operational disruption. This means that multi‑Agent systems must address not only “how to divide tasks” but also “how power is allocated.”

Returning to that retail company, a single store transfer must pass through several checkpoints.

The replenishment Agent of Store A has its own identity, and the system knows which store it represents and what role it plays. It can only read and write data for its own store and cannot see the complete business information of other stores. It can send a transfer request to the Agent of Store B, and Store B decides how much to release based on inventory; once the transfer amount exceeds the set limit, the process pauses, awaiting regional manager confirmation.

“What I see when I come in the morning are already negotiated orders; I’ll approve the large ones,” said the regional operations manager.

The hardest part of this mechanism is permissions.

The Agent interacts with many systems, and each call could become a risk entry point. Therefore, keys should not be exposed to the Agent long‑term, but should be issued on demand at runtime to the appropriate tools and clients; before the actual call, the identity, task, and permissions should be re‑checked for a complete match.

Large abnormal operations are usually easy to detect, but a more covert risk is a large number of small transactions. For example, if an Agent gradually transfers assets through many small transfers, the system must not only check each individual call but also identify anomalous patterns from historical traces.

Model agents, tool access, and external system connections therefore need to be brought under unified control so that the authority scattered across different Agents can be consolidated, audited, and revoked.

So, who is listening to whom among the 5,000 concurrent “Einsteins”?

The answer is not to establish a permanent central brain, but to assign each Agent a clear identity and context, allowing collaboration to occur in individual “rooms”.

Members in the room share the context needed to complete the task. The Leader Agent is responsible for breaking down the task: deciding which work to assign to team members, which to handle personally, and how to consolidate the results. When a step requires human responsibility or judgment, the process pauses and hands control back to a person; every prior operation is logged for later audit.

“Once you leave this room, this information may no longer be relevant to others.” When explaining this concept, Zhou Qi gave an analogy: “It’s like the discussion we’re having today.”

Those Remaining Traces

The room will disappear, but the trace will remain.

“When building an enterprise system, the timeline is usually measured in years. A design from a year ago might cover 80% of scenarios, while the remaining 20% of long-tail requirements may remain unaddressed for a long time due to excessive investment. Companies often have to wait for the next budgeting cycle before they can rebuild the system.”

The generalization ability of AI makes it possible for these long-tail needs to be captured for the first time, and the process of an Agent completing a task can itself become an enterprise asset.

In his early career, Zhou Qi worked on log processing and system monitoring diagnostics. Here, the recorded objects extend further into the work process: not only identifying what went wrong with the system, but also tracking which attempts were made, which tools were invoked, and why a task ultimately succeeded or failed.

In AgentCore, the component responsible for collecting these trajectories is called AgentLoop. It records the attempts made by a human or an Agent, the tools invoked, and the final outcomes, which are then used for subsequent organization, evaluation, and reuse.

A successful experience can be distilled into an SOP, a tool, a Skill, or a memory. When a similar request arises again, the system can directly invoke the existing experience instead of planning from scratch.

The operational troubleshooting at Manbang has comparable value. Known fault localization and resolution processes can be captured as corresponding Skills and knowledge base entries. The next time a similar issue occurs, it won’t require a cold start due to employee forgetfulness or team turnover.

In the past, people’s experience was often scattered across tickets, chat logs, and personal memory. After an employee leaves or changes roles, a company may only retain the outcome, without knowing why that judgment was made at the time. The Agent execution trace gives enterprises their first chance to systematically record “how the work was done.”

But experience only has reuse value when it can be found.

Another component is Agent asset management, which is responsible for centrally managing Skills, MCP tools, models, and Agent templates, recording different versions, and tracking their usage, ratings, and evaluation results. In an ideal state, every new employee or Agent doesn’t have to start from scratch but can “stand on the shoulders of others.”

Databases are just one storage medium for these assets. Customer service records, shipping orders, client preferences, and the data accumulated by a company through long‑term digital transformation can all serve as context for subsequent tasks. Companies that have already completed digitalization can also leverage existing data to cold‑start AI projects.

But the fact that a trajectory has been recorded does not mean it will remain valid forever.

After model version upgrades, tool changes, or adjustments to task objectives, previously successful paths may become obsolete. Enterprises need to continuously perform regression evaluations on historical cases to verify whether the new version’s performance has improved or declined.

“You can’t bet everything on probability,” Zhou Qi said.

This type of assessment is a bit like giving employees a yearly rating: whether goals were met, how many resources the process consumed, whether quality standards were achieved, and whether any authority boundaries were crossed.

Costs must also be calculated in this process. Different tasks can select different models through intelligent routing; when an Agent is inactive, the system can automatically scale down to reduce idle resources. Tokens, fees, latency, and success rates generated by each call also need to be observed separately for the caller, the Agent, and the model.

Therefore, building an Agent is only the first step; the more challenging part is continuously operating it in the future.

“Enterprises need to preserve effective methods, re‑evaluate them when models and tasks change, and then adjust identities, tools, processes, and collaboration relationships based on the results,” says Zhou Qi. Traditional software architectures may change on a quarterly or yearly cycle, whereas the structure of an Agent system could theoretically evolve every day.

Where does a company start?

Zhou Qi’s advice is simple: first find a relatively certain, high‑frequency, easily digitized scenario with measurable results, and run the full workflow on a unified platform; then gradually bring more people and teams from upstream and downstream into the process, turning the task from a single point into complex collaboration.

Finally, return to the organization.

Zhou Qi says that both individuals and companies should aim to become “friends of time” in the future, continuously building assets that generate compounding returns. These assets include business knowledge, task trajectories, tacit experience, as well as the collaborative patterns formed between people, between people and agents, and among agents.

The earlier you start, the more pronounced the compounding benefits become.

Executable tasks can be gradually handed over to the Agent, while humans take on more of the division of labor, judgment, and review.

Reporter asks: In such an organization, what is the most important value of humans?

“Review, define the standards well.”

END

Share, like, view, schedule?

Today’s Recommended Articles

Anthropic team starts with zero Rust experience, using Claude to complete core API cross-language refactor: 40% CPU savings, 20% latency reduction

Anthropic eliminates coding interviews: newcomers get a chance to compete directly with senior engineers on their first try

Claude Code supports the largest IPO in history, Chinese developers whose accounts were banned without refunds support that 5% “dirty income”

Microsoft invites 60,000 employees to voluntarily resign: covering over a quarter of its global staff, even vice presidents are included