Alibaba Cloud AgentCore: Multi-Agent Collaboration Architecture for Complex Task Automation

Enterprises often face cross-module tickets that require multiple professional capabilities: interpreting business rules, verifying system data, analyzing root causes, and synthesizing conclusions into actionable responses. While organizations have accumulated business documents, system interfaces, and operational standards, the challenge lies in organizing these capabilities so that specialized agents can collaborate with clear division of labor, visible processes, traceable results, and optimizable outcomes.

01 Why Complex Tasks Need Professional Division

When a single agent's scope expands, its tools, knowledge, and rules grow accordingly, creating three engineering pressures:

Tool and skill overload: Overlapping tool and skill applicability forces the agent to identify the right capability, invocation method, and permissions for each task.

Cross-domain context pollution: Business knowledge, history, and intermediate results crowd the same context, increasing the burden of filtering key facts, maintaining constraints, and validating conclusions.

Centralized maintenance impact: Shared prompts, knowledge, and tool configurations mean a local change can ripple across unrelated responsibilities, requiring broader validation.

Professional division addresses these issues by assigning domain knowledge and tool configurations to dedicated agents with explicit inputs, responsibilities, and delivery requirements. For example, a ticket-answering system can use a domain agent for rule interpretation, a verification agent for data queries, and a diagnostic agent for system analysis. Work proceeds in parallel where possible and sequentially where dependent, with a leader agent consolidating the final response. Boundaries should align with business capabilities and maintenance ownership: simple, well-bounded tasks suit a single agent; multi-domain, continuous, or acceptance-gated work fits a team.

02 Humans and Agents Working in a Shared Collaboration Space

AgentCore's collaboration architecture centers on four pillars: AgentIdentity for identity mapping, Matrix for room-based messaging, Team and Task for organization and delivery, and observability/governance for analysis and optimization.

▍ Every Person and Agent Has an Identifiable Identity

Human members, managed agents, and custom agents each participate with independent identities. AgentIdentity integrates existing account systems and maps them to Matrix identities. This clarifies who requested work, which agent executed it, and who submitted results. Resource access remains role- and permission-based; joining a collaboration room does not grant blanket data or tool access.

▍ IM-Style Communication Makes Collaboration Visible

Matrix extends human messaging patterns to human-agent and agent-agent interactions. Members post plans, progress updates, supplementary information, and result references in rooms, creating a timeline that anyone can follow. Visible information includes public analyses, execution progress, and deliverables. Human experts can inject judgments, correct data, or participate in acceptance within the same space, keeping collaboration close to daily communication habits.

▍ Clear Team Roles Ensure Professional Accountability

Team admins (human) configure team metadata, members, and the leader agent. The leader agent understands goals, plans, coordinates division of labor, and aggregates results. Worker agents execute domain-specific tasks, while human tech leads or domain experts join for judgment and acceptance. Enterprises retain each business team's maintenance responsibility while enabling cross-functional collaboration on shared tasks.

03 Task System Drives Work to Completion

Identity and communication solve participation and dialogue. A business job also needs clear goals, ownership, progress tracking, and acceptance criteria. AgentCore defines a task as work with explicit objectives, responsibility assignments, execution states, and delivery requirements, organized through three core objects:

Task: The overall job, recording business goals, execution team, and acceptance criteria.

SubTask: Professional assignments specifying who owns which part, interdependencies, and execution status.

Result: Delivery artifacts storing actual outputs, result versions, and review comments for verifiable acceptance.

In the cross-domain ticket example, the leader decomposes the goal into subtasks for rule lookup, data verification, and problem analysis. Members submit artifacts; the leader advances work based on dependencies and completion, then forwards the consolidated result to configured reviewers for acceptance. During execution, users can view progress, add information, inspect subtasks and result files. Blockers or change requests are recorded continuously, allowing the team to resume from the latest state. Operations follow task status and permissions.

Mechanically, TeamHarness provides role-adapted collaboration skills for leaders and workers, while agents interact with the task service via a dedicated CLI. The leader side handles planning and task management; the worker side handles task acceptance, status reporting, and result submission. The task service validates permissions and state transitions; Matrix handles messaging and notifications. This lets business teams focus on domain responsibilities, assets, and delivery requirements while the unified task model captures division of labor, states, events, and results. Details are in the TeamHarness Usage Guide (https://help.aliyun.com/zh/agentcore/team-collaboration-instructions-collaboration-teamharness) and Team Task Documentation (https://help.aliyun.com/zh/agentcore/using-team-tasks).

For recurring work, enterprises can configure scheduled tasks via UI or conversational agent interaction. Each trigger creates an independent execution record tracked by the same task mechanism. Disabling the schedule stops future triggers; canceling a single run stops only that execution.

04 Building an Enterprise Multi-Agent Team

Adoption follows a path: prepare assets, configure professional agents, form teams, connect business entry points.

▍ Organize Enterprise Accumulation into Usable Assets

Knowledge bases and retrieval services provide business grounding. Curate documents, domain knowledge, rules, and historical cases into searchable repositories with clear ownership, scope, and update processes.

MCP and CLI toolsets expose operational capabilities. Existing business APIs can be published as MCP tools via an AI gateway with descriptions, parameters, and permissions. Existing command-line capabilities become CLI tools with prepared dependencies, execution environments, and access controls.

Skills codify reusable work methods. Transform SOPs and domain experience into structured guides covering applicability conditions, processing steps, exception handling, and delivery requirements, directing how agents apply knowledge and tools.

These assets are independently maintainable and flexibly combined per agent role. A single knowledge service or business tool can be reused by multiple agents within authorization scopes.

▍ Configure Capabilities and Models by Business Role

For each professional agent, define responsibilities, inputs, and delivery boundaries, then assign matching knowledge, skills, tools, models, and permissions. Organizational division clarifies maintenance ownership; granularity should follow business capabilities and delivery interfaces.

Models can also be role-specific. Complex planning and cross-domain synthesis demand different capabilities than rule-based extraction and classification. Enterprises can select appropriate models per role under quality constraints, comparing cost and effectiveness on real task data.

▍ Form Teams and Let Users Initiate Collaboration via Familiar Entry Points

Once professional agents operate independently, team admins configure leaders and members, specify human expert involvement, and assemble capabilities into business-oriented teams. AgentCore offers five access channels:

Web collaboration client

DingTalk

Feishu (Lark)

WeCom (Enterprise WeChat)

AG-UI protocol for custom frontends or self-built applications

The three enterprise IMs connect via leader-associated bots; business frontends use AG-UI. Users submit goals through familiar interfaces; the leader coordinates specialists and returns results. Channel configurations and presentation differ; enterprises choose based on work scenarios.

05 AgentLoop Enables Continuous Iteration in Production

As multiple agents run concurrent tasks, execution chains, dependencies, and call logs grow. Enterprises need to understand from the business task perspective: which step is waiting, which tool fails, which calls cause redundant consumption.

AgentCore supplies task and execution data; AgentLoop provides observability, evaluation, and continuous optimization. Linking Task, SubTask, and Agent Trace enables analysis of collaboration performance through task events, model and tool invocations, and delivery outcomes. Correlation analysis requires corresponding data ingestion and identification.

Enterprises can structure continuous iteration in four steps:

Collect execution data: Correlate tasks with agent traces, retaining invocation logs, deliverables, and business feedback.

Evaluate collaboration effectiveness: Curate typical and failure samples, combine evaluation results with manual review to pinpoint issues.

Adjust collaboration configuration: Revise knowledge, prompts, and skills; tweak professional division, task dependencies, model and tool selections.

Regression validation: Compare versions on the same task set, confirm improvements against business standards, then monitor new execution data.

Evaluation focuses on three dimensions: delivery quality (accuracy, evidence completeness, rework rate), delivery cycle (actual execution vs. dependency wait time), and unit delivery cost (model, tool, retry, and human investment). Every configuration change thus has comparable evidence; validated assets and configurations remain reusable. AgentLoop's observability and evaluation capabilities are detailed in the product overview (https://help.aliyun.com/zh/agentloop/product-overview/what-is-agentloop).

Start with a Clear Business Goal

Cross-domain ticket answering is a practical entry point for organizing professional agents. Enterprises can begin with a work category that has clear responsibility boundaries, available knowledge and interfaces, and verifiable outcomes. Define acceptance standards, then assemble the required assets and members. AgentCore embeds professional division into identity, communication, team, and task mechanisms, allowing enterprises to organize existing knowledge and system capabilities, continuously track work, and measure collaboration effectiveness by business results.

References

[1] TeamHarness Usage Guide: https://help.aliyun.com/zh/agentcore/team-collaboration-instructions-collaboration-teamharness

[2] Team Task Documentation: https://help.aliyun.com/zh/agentcore/using-team-tasks

[3] AgentLoop Product Overview: https://help.aliyun.com/zh/agentloop/product-overview/what-is-agentloop

[4] AgentCore Official Website: https://www.aliyun.com/product/agentcore

[5] Team Task Usage Documentation: https://help.aliyun.com/zh/agentcore/using-team-tasks