Alibaba Cloud AgentCore: Enterprise Agent Platform for Collaboration, Governance & Scale

Over the past year, AI agent development has accelerated with numerous production deployments and new technologies. The article identifies two major enterprise agent scenarios: functional role intelligence (personal desktop assistants, digital employees, software development automation) and business scenario intelligence (intelligent customer service, logistics, marketing, audio/video). For functional roles, agents target all employees including non-technical staff, requiring low-barrier creation and desktop-cloud collaboration, with success measured by per-person efficiency gains. For business scenarios, agents require deep business personnel involvement, multi-agent orchestration, security/compliance/isolation, and validation against direct business metrics like customer satisfaction and human intervention rates.

Agent Forms and Evolution

Three agent forms have converged in 2024: open-source harness frameworks (e.g., OpenClaw) for technically strong enterprises; commercial general agents (e.g., Alibaba's Qwen Office) for functional roles; and high-code custom agents for complex business logic. Workflow-based agents, popular in 2023, have converged to high-determinism, high-compliance scenarios. The core challenge has shifted from agent construction (now lower barrier) to scaling business results through collaboration, asset沉淀 (immersion), evolution, and governance.

AgentCore Platform Overview

AgentCore is a one-stop enterprise-grade agent construction and governance platform. It supports multiple build modes: high-code (any SDK like LangChain, AgentScope, plus AgentCore SDK for platform assets), no-code managed harness (Qwen-Paw, Qoder CLI), and governance of external agents (desktop, local). Deployment follows standard software pipelines (code, image, CI/CD). The platform provides fully managed runtime with isolated sandboxes coupled to agent lifecycles, elastic scaling, version management, and canary releases. Integration uses standard AG-UI and A2A protocols plus IM/DingTalk connectors.

Multi-Agent Collaboration Patterns

Three collaboration patterns are identified: (1) Instant concurrent — a main agent spawns parallel sub-agents sharing context/permissions (e.g., market research), natively supported in harness frameworks. (2) On-demand teaming — predefined agent templates execute sequential steps (research → write → review) with lifecycles bound to the task. (3) Long-term digital employees — persistent role-based agents (e.g., product manager) with permissions, observability, and evolution, invoked as needed for main/sub-agent coordination. AgentCore implements a Team mode placing humans and agents as peers, with a Leader Agent coordinating Worker Agents and a chat-like digital meeting room providing transparency, visibility, and real-time human intervention via @mentions.

Context Engineering: Memory and Knowledge Fusion

AgentCore integrates Alibaba Cloud knowledge bases (Bailian, OTS) and memory services (OTS Memory, PolarDB Memory) via MCP, plus a built-in memory service. The boundary between memory and knowledge is blurring: long-term memory extracted from conversation histories can be沉淀 (precipitated) into knowledge bases and Skills. Planned capabilities include automatic precipitation of high-value online executions into knowledge/Skills, unified hybrid retrieval across memory and knowledge, and permission isolation for shared vs. isolated contexts in multi-agent collaboration.

Enterprise Asset Management

Four core asset types: Skills (becoming core business logic carriers, replacing code for many workflows), MCPs (tool connections), model connections, and agent templates (best-practice bundles). Skills require private enterprise Skill markets with versioning (via built-in Nacos), security scanning (static code, dependency vulnerabilities, prompt injection), distribution controls, and evaluation/tuning. MCPs need market management, RESTful-to-MCP conversion, and permission control. Model connections require multi-vendor management, per-department/agent model lists, and runtime smart routing, fallback, rate limiting, and quotas. Agent templates enable one-click cloning.

Three-Layer Catalog Architecture

Personal layer: employees create private agents/Skills. Enterprise layer: approved assets shared company-wide (department Skills, excellent personal contributions). External layer: curated third-party assets (Alibaba Cloud official agents/Skills). AgentCore implements personal and enterprise catalogs with permission isolation, metadata management, access control, onboarding standards, and operations analytics (usage, callers, SLA).

Observability and Continuous Optimization (Data Flywheel)

AgentCore integrates AgentLoop observability. The data flywheel: (1) Collection — capture every step (model calls, tool invocations, knowledge retrieval). (2) Analysis — efficiency metrics (rounds, tool counts), token cost dashboards, performance breakdowns (model vs. tool latency). (3) Evaluation — due to LLM non-determinism, deterministic rules fail; three evaluator types: LLM-as-a-Judge (semantic relevance), Code-as-a-Judge (deterministic rules), Agent-as-a-Judge (business-rule-heavy, using private knowledge/tools/expert rules). Customer practices show Agent-as-a-Judge achieves >90% consistency with expert human labels, replacing manual annotation at scale. AgentCore supports AI-assisted evaluator creation via natural language. (4) Optimization — findings feed back into agent improvements.

Governance: Identity, Security, Cost

Every agent gets a unique identity via Alibaba Cloud Agent Identity. Inbound access control integrates with enterprise IdPs (DingTalk, Lark, WeCom, custom). Outbound control via built-in AI gateway defines per-user/per-agent MCP and external system permissions. Full tracing enables auditability and root-cause analysis. Cost governance: observability provides per-request, per-agent, per-user token accounting; alerts prevent month-end surprises; AI gateway enforces quotas with throttle/degrade/block actions.

Conclusion

AgentCore positions as an enterprise-grade agent construction and governance platform. Infrastructure enables rapid building; collaboration, asset沉淀, evolution, and governance enable scaled business outcomes. A demo and DingTalk group (185130001370) are available for further engagement.