Scaling Multi-Agent AI Across 4,360 Retail Stores: Top Sports' Enterprise Practice

Top Sports operates 4,360 physical stores, 3,700+ mini-program stores, partners with 20+ sports brands, and serves 90M+ users. The challenge: expertise cannot rely on individual experts being present at every store. Different roles need different help — frontline staff ask about product pitches and training, while store managers need diagnostic analysis and actionable adjustments.

Why Multiple Enterprise Agents Are Needed

Goals across roles are related but required knowledge, data, and permissions differ. Therefore, distinct agents are needed for distinct positions. However, building multiple agents introduces three core problems:

Duplicate construction: Same knowledge, data, and tools re-integrated per project.

Collaboration and permissions: How do operations, membership, and marketing agents hand off work? What can each role see and do?

Operations accountability: How to measure effect, trace errors, and assign ongoing ownership after launch?

These cannot be solved per project; a unified operational foundation is required. AgentCore provides that foundation: unified management of model connections, MCP tools, Skills, and Agents, plus Workspace, team, role, and credential management for clear resource boundaries. Agent Team capability lets multiple specialized agents collaborate around a task. Issues become traceable, and combined with Alibaba Cloud's AgentLoop for evaluation and improvement, the system avoids "launch and forget."

Platform Architecture: Bottom-Up Reusability

The stack layers: knowledge, business data, MCP and API tools at the base; reusable Skills (product query, inventory, member segmentation) above; professional Agents (Operations, Product, Training, Member, Marketing) on top; Agents form teams around business goals. Users interact via an AI Assistant; a Team Leader decomposes tasks, schedules capabilities, and aggregates results. Identity permissions, runtime observability, and evaluation feedback run throughout.

Three Concrete Scenarios

1. Frontline Product Training (练货)

Operations Agent detects sales deviation for a product. Deviation ≠ training gap, so field feedback confirms real need. Product Agent supplies product details, selling points, pairings. Training Agent designs interactive practice sessions. Scores and suggestions feed back to AI Assistant for scheduling remedial training and ongoing observation. Monthly stats: 5,700+ products involved, 390,000+ training sessions. Key question: does training translate to actual service improvement?

2. Lead Mining for Online Store Managers

Online mini-program store has a single manager handling data, customer selection, product picking, and communication. Generating a report adds reading load without reducing work. Instead, Operations + Member Agents identify priority customers; Product Agent matches preferences and available stock; Marketing Agent prepares communication scripts; AI Assistant consolidates into executable daily tasks. Manager confirms and executes. Mini-program revisit and conversion records feed next cycle.

3. Store Operation Planning

Manager states goal and constraints (e.g., weekend plan). No preset button. AI Assistant calls capabilities to assemble performance, inventory, member data into recommendations: product pushes, display adjustments, customer maintenance. Manager iterates, modifies, confirms. Critical: capability reuse ≠ data sharing — manager accesses only own-store authorized data. Result: planning time reduced from 2 hours to ~20 minutes. Freed time shifts to business judgment.

From Scenarios to Reusable Capabilities: Four-Step沉淀 Process

Select scenario: high-frequency, rigid-demand, measurable problems; define original process duration and acceptance criteria.

Build with沉淀: capture knowledge, tools, Skills, Agents, and evaluation data during construction for reuse.

Joint operations: Business, IT, and AI teams jointly operate post-launch, continuously monitoring usage and value.

Optimize or retire: effective capabilities are optimized and replicated; long-term low-value scenarios are adjusted or exited.

After each scenario, ask: (1) What did we solve? (2) What can be reused next time? Verified handling methods become SOP-type Skills with explicit applicability conditions, input processing, tool calls, outputs, normal/exception flows, and human-confirmation points. Skills are versioned, owned, and evaluated. Stable Skills compose into professional Agents — leaving not just a one-off answer but a verified solution pattern.

Shared Responsibility Model

Business knows problems and judgment rules → defines acceptance criteria, owns scenarios. IT owns platform, integration, identity, stability. AI team owns agent engineering, common capabilities, evaluation, version optimization. Division of labor, but not divided acceptance. Post-launch, all three decide together whether to optimize or retire based on shared business outcomes — avoiding "Agent launched, responsibility gone."

Continuous Tuning: Dual Loops

AgentLoop (capability improvement) and Business Loop (business problem solving) are two sides of the same coin. Metrics tracked:

Adoption: calls, but also acceptance rate of suggestions.

Quality: correctness, exceptions, human fallback rate.

Efficiency: duration, time saved, per-call cost.

Business: execution rate, outcome improvement.

Dashboard shows what happened; operations answer "what to fix next?" — adjust entry points, repair tools, update Skills. Improvements enter next cycle. Example: if training effect poor, examine training needs, content, evaluation. If customer maintenance misses target, analyze audience selection, product matching, or communication content. Real feedback drives Skill/Agent updates; key decisions and execution remain human-owned.

Conclusion

The goal is not more agents, but turning today's solved problems into tomorrow's reusable capabilities. AgentCore provides the build-and-governance foundation; AgentLoop provides the data flywheel; Top Sports continuously沉淀s retail experience into reusable, operable enterprise assets.