Probability, Sockets & 5000 Einsteins: Alibaba Cloud VP on Enterprise AI Agent Infrastructure

The Core Contradiction: Probability vs. Determinism

Zhou Qi, VP of R&D at Alibaba Cloud, opens with a historical analogy: late‑19th‑century factories ran on a central steam engine driving a long transmission shaft; early electric motors merely replaced the steam engine while keeping the shaft and belts. Economic historian Paul David (1990) showed that electric motors accounted for under 5% of factory mechanical power as late as 1899, and productivity gains only appeared in the 1920s after factories redesigned around distributed electric drive. Zhou argues that today’s enterprises have their own “transmission shaft” — decades of code, processes, and management built on formal logic (deterministic inputs, conditions, outputs) — while AI’s foundation is probability. This creates a fundamental tension: personal AI tolerates pixel‑level errors, but enterprise AI must deliver ROI over years and therefore requires certainty.

From “Having Electricity” to “Having Sockets”

When electricity entered homes, wiring was only for lighting. To use an iron or fan, users unscrewed the bulb and plugged into the lamp socket. Harvey Hubbell’s 1904 separable plug patent helped, but the base still screwed into the lamp holder. The real problem shifted from “is there electricity?” to “how do I connect?” By 1915‑1919 the U.S. market had 15‑20 plug types and 15‑30 socket types, none interchangeable — a standardization crisis mirrored in today’s agent landscape. Without a common “socket,” each team builds its own workshop: wiring models, writing tools, configuring permissions in isolation. Agents cannot identify each other, inherit permissions, or interoperate, producing many demos but few production deployments.

Industry focus has moved from “can we build an agent?” to “how do long tasks run continuously, how is state recovered, how are credentials managed, how is behavior audited?” The runtime and governance infrastructure for agents is becoming a shared enterprise need.

AgentCore: A Standardized “Power Strip” for Enterprise Agents

Alibaba Cloud’s open‑source AI Infra components correspond to emerging socket standards; AgentCore is a one‑stop enterprise‑grade agent building and governance platform. It handles the “wiring, transformers, and circuit breakers” so teams can focus on business logic. Crucially, standardization does not discard legacy systems — accumulated processes, APIs, and data become the “hands and feet” that let AI act. As Zhou puts it, “without hands and feet, AI cannot fully deliver value.”

Case Study: Manbang’s Stability “Digital Human”

Context: Manbang connects shippers and truckers; 2025 fulfilled orders reached 2.36 billion (+19.8% YoY). A single dispatch crosses the truck‑side app, driver app, shipper app, and backend scheduling system — any network, database, or service hiccup fails the task.

Pain point: Ops engineers monitored 5 systems and opened 6 platforms during incidents. “Fixing takes five minutes; finding the root cause takes half an hour.”

Solution steps:

MCP‑ify existing ops tools into standard interfaces callable by agents.

Bring the existing RBAC/approval system into the agent runtime (who can take down a machine, restart an app, change config, which faults are auto‑remediated, which need human sign‑off).

Result: The stability digital human can execute thousands of commands per minute. At the 2026 Yunqi Conference Manbang disclosed that in major‑risk scenarios the agent locates the cause within 1 minute with >90% accuracy .

Human impact: Engineers shift from nocturnal firefighting to higher‑value work.

Four Pillars the Infrastructure Must Solve

Stability: Not just uptime, but keeping AI on the enterprise’s intended path — translating industry‑specific abstractions into rules, boundaries, and evaluation criteria.

Security: Protecting the enterprise’s existing foundations: roles, permissions, compliance, privacy. Traditional software controls deterministic calls; agents autonomously plan and chain tools, expanding the governance surface to identity, permissions, data, tools, memory, behavior, and runtime environment.

Performance & Cost: Long‑running tasks (days) must survive machine failures without state loss; recovery points, data access, and operation rights are managed uniformly. Intelligent routing picks models per task; idle agents auto‑scale down. Every invocation’s tokens, cost, latency, and success rate are observed per caller, agent, and model.

AI‑Native Planning vs. Traditional Workflows

Traditional software defines a fixed workflow: step 1, step 2, exception branch. AI‑native planning resembles cooking: prepare ingredients, declare dependencies and goal, set boundaries, let the agent dynamically determine steps. A year ago enterprises lacked confidence in long tasks and used AI only for small tools; now they experiment with goal‑only definitions. “The latter has higher uncertainty but a higher ceiling,” Zhou notes.

Multi‑Agent “Rooms”: 5000 Einsteins Collaborating

Zhou envisions a complex task instantly handled by 5000 top‑tier reasoning agents (“5000 Einsteins”) — some researching, some coding, some marketing — then synthesizing a plan. The challenge: who listens to whom? Who sees what data? Who decides on disagreements?

A retail chain illustrates the “room” model: each store gets an agent team (replenishment, reconciliation, coordination agents). Store‑A’s replenishment agent has its own identity, reads/writes only its own data, requests stock from Store‑B’s agent, which decides based on its inventory. Transfers above a threshold pause for regional manager approval. “Now morning shows only pre‑negotiated deals; I only approve the big ones,” says the regional manager.

Security risks grow with collaboration chains. A Cloud Security Alliance survey found 65% of organizations experienced at least one AI‑agent‑related incident in the past year (data exposure, operational disruption). Keys must not be long‑lived; they are issued at runtime to specific tools/clients, with identity‑task‑permission checks before each call. Systems must also detect “low‑and‑slow” attacks — many small transfers that evade single‑call thresholds — requiring a unified control plane to collect, audit, and revoke scattered permissions.

The answer to “who commands 5000 Einsteins?” is not a permanent central brain but ephemeral “rooms” where members share only the context needed for the task. A leader agent decomposes work, assigns subtasks, aggregates results, and pauses for human judgment when responsibility or discretion is required. Every step leaves an audit trail. “Once you leave the room, that information may be irrelevant to others,” Zhou says.

Traces That Remain: AgentLoop and Asset Management

Enterprise systems traditionally take years to build; a year‑old design covers ~80% of scenarios, leaving a long tail of unaddressed needs. AI’s generalization lets agents catch those long‑tail cases, and the agent’s execution trace itself becomes an asset. Zhou’s background in log processing and monitoring extends to recording how work gets done : attempts, tool calls, success/failure reasons.

AgentLoop captures human/agent attempts, tool invocations, and outcomes for later curation, evaluation, and reuse.

Successful experiences become SOPs, Skills, tools, or memories. On recurrence, the system reuses proven paths instead of replanning.

Agent Asset Management governs Skills, MCP tools, models, and agent templates with versioning, invocation counts, ratings, and assessment results — enabling new hires (human or agent) to “stand on shoulders.”

Existing databases, service logs, shipping records, customer preferences — all become context for future tasks. Enterprises with mature digitization can cold‑start AI projects from this stock.

But assets decay: model upgrades, tool changes, goal shifts invalidate old paths. Continuous regression evaluation (like annual performance reviews) checks goal completion, resource consumption, quality, and boundary violations. “You can’t bet everything on probability,” Zhou warns.

Starting Point and Organizational Evolution

Zhou’s advice: pick one relatively deterministic, high‑frequency, easily digitized, measurable scenario; run the full loop on a unified platform; then gradually onboard upstream/downstream teams to expand from point solutions to complex collaboration.

Ultimately, both individuals and enterprises must become “friends of time,” continuously accumulating compounding assets: business knowledge, task traces, tacit experience, and collaboration patterns among humans and agents. The earlier you start, the greater the compounding effect. Execution work shifts to agents; humans focus on division of labor, judgment, and review. When asked what human value remains, Zhou answers: “Review, define good standards.”