TRACK DETAILS

Track Details

Explore the four tracks, challenge requirements and participant support.

Global Open Source RMB 5M Prize Pool

Agent Infra

Overview

Focus on multi-agent infrastructure and collaboration systems for complex enterprise tasks, moving agents from demo to production.

This track focuses on multi-agent infrastructure and collaboration systems for complex enterprise tasks. Teams are encouraged to start from industries, professional domains, or real business scenarios they know well, identify complex tasks with broad industry relevance, and build Agent Infra solutions with reusable Skills, tool integration, and runtime verification capabilities.

The track is not about showing what a single agent can do. It asks how multiple agents complete task decomposition, context passing, tool calling, result verification, execution-evidence capture, and security auditing under complex constraints.

Entries should target enterprise application scenarios and design a complete closed-loop solution with at least three agents in different roles, forming an end-to-end task loop for real enterprise settings.

Prize Information

Grand Award1 team / RMB 1M

*The winner of the Grand Award will be selected from the four track champion teams.

Champion1 team / RMB 500K
Runner-up1 team / RMB 300K
Third Prize1 team / RMB 100K

Additional Awards

Open-Source Rising Star Award4th-10th place, gifts worth RMB 5,000 per team.
Open-Source Impact Award11th-15th place, gifts worth RMB 3,000 per team.
Excellence AwardSemi-finalist teams receive a commemorative gift package worth RMB 500.
Participation AwardTeams submitting valid entries receive a competition T-shirt.

Topic Directions (Open-ended Topics)

Teams may use the following directions as references, but they are scenario prompts only and do not limit the eligible scope. The organizer encourages participants to start from their professional background, industry, research practice, enterprise projects, or real work and learning scenarios, discover complex tasks with broad industry relevance, and propose solutions in domains where they have stronger insight.

Entries do not need to be limited to operations, security, or cost governance. Any topic from real demand that shows the value of multi-agent task decomposition, collaborative execution, tool calling, result verification, and experience capture, and that can potentially become reusable Skills, tool interfaces, or open-source capabilities, is valid for this track.

The following directions can help teams understand the track positioning and ideate solutions:

Direction 1: Zero-touch Operations

Build a multi-agent loop around alert aggregation, root-cause localization, remediation execution, recovery verification, and incident review.

Reference scenarios include:
  • Multi-source alert aggregation and noise reduction
  • Automated root-cause localization
  • Remediation recommendation and execution
  • Service recovery verification
  • Incident review and knowledge capture
Direction 2: Autonomous Customer-Service Loop

Build a multi-agent loop around multi-channel conversation aggregation, intent recognition and prioritization, solution generation and execution, result verification and customer confirmation, and case review with knowledge capture.

Reference scenarios include:
  • Multi-channel ticket / conversation aggregation and deduplication across email, online support, phone transcripts, and social comments
  • Intent recognition and automatic ticket classification and prioritization
  • Solution generation and automated execution, such as refunds, exchanges, or account changes
  • Result verification and customer-satisfaction confirmation
  • Difficult-case review and support-knowledge-base capture
Direction 3: End-to-end Software Development Collaboration

Build a multi-agent loop around defect / requirement aggregation, code root-cause localization, fix generation and execution, test verification and release confirmation, review and knowledge capture.

Reference scenarios include:
  • Multi-source defects / requirement aggregation and deduplication across issues, logs, and user feedback
  • Automated code-defect localization and impact analysis
  • Fix generation and automated coding execution
  • Test verification and canary-release result confirmation
  • Post-release review and development-knowledge-base capture
Direction 4: Financial Risk Control and Claims Automation

Build a multi-agent loop around multi-source risk-signal aggregation, risk localization and verification, response-plan generation and execution, result verification and compliance audit, incident review and risk-knowledge capture.

Reference scenarios include:
  • Multi-source risk-signal aggregation and noise reduction from transaction records, credit data, public sentiment, and complaint records
  • Automated risk / fraud root-cause localization and impact analysis
  • Claims or credit response recommendation and automated execution
  • Response-result verification and compliance-audit confirmation
  • Risk-event review and risk-control knowledge-base capture

Technical Requirements

This chapter contains cross-stage technical requirements. The preliminary round may focus on design thinking, while the semi-final and final should gradually form runnable and verifiable engineering materials.

01 Multi-agent Collaboration Requirements

1.1 Multi-agent Collaboration Requirements

The solution should design at least three agents with different roles. Each agent needs a clear identity definition and should collaborate to complete an end-to-end task loop. Multi-agent design must use AgentTeams as the collaboration design baseline and explain how role orchestration, task decomposition, context passing, collaborative execution, and state tracking map to the framework capability.

1.2 Agent Identity List

Teams need to submit an Agent Identity list explaining each agent identity, capability boundary, and collaboration relationship in the multi-agent system. See Appendix A of the Participant Handbook.

1.3 Multi-agent Closed-loop Description

The solution should explain how multiple agents complete the following loop:

  • Task input: the system receives alerts, tickets, logs, bills, security events, or other enterprise task inputs.
  • Task decomposition: the controller agent or collaboration framework decomposes the task for agents with different roles.
  • Context passing: agents pass task context, history, tool-call results, and intermediate conclusions.
  • Tool calling: agents invoke Skills, MCP tools, cloud products, enterprise systems, knowledge bases, or external APIs.
  • Result verification: the system verifies remediation, security response, cost optimization, or other results.
  • Execution-evidence capture: the system stores logs, Trace, Metrics, reports, screenshots, or other execution evidence.
  • Approval and rollback: high-risk actions need human confirmation, approval, rollback, and audit mechanisms.
  • Experience capture: execution results, reviews, or experience rules are packaged as reusable capabilities.

02 Skill, MCP, RAG, and Observability Requirements

2.1 Skill

Skill is mandatory for this track. Teams may use official Alibaba Cloud Skills or package key capabilities as reusable Skills. Skill should act as the task-capability abstraction layer rather than a one-off agent behavior. Each solution should provide a core Skill list and explain:

  • Skill name
  • Skill purpose
  • Inputs and outputs
  • Invocation conditions
  • Dependent tools
  • Failure handling
  • Security boundary
  • Reuse value
  • Relationship to the multi-agent collaboration flow
2.2 MCP and Tool Integration

MCP is recommended. It is the recommended protocol for connecting external tools and systems such as cloud products, enterprise systems, databases, knowledge bases, ticketing systems, CI/CD, and monitoring systems. Skill acts as the task-capability abstraction layer, while MCP acts as the tool-connection layer.

  • If MCP is not used, the solution should provide an equivalent external-tool integration contract covering protocol, authentication, input / output schema, error handling, audit records, and the cost of later migration to MCP.
  • The equivalent contract should abstract the external tool into a stable capability callable by Agent / Skill and explain tool name, invocation entry, parameter schema, response structure, permission scope, failure retry, idempotency control, audit logs, and fallback behavior.
  • Reviewers do not require an MCP Server implementation for alternatives, but should be able to judge whether later migration to MCP only requires protocol adaptation rather than redesigning the tool-call chain.
2.3 Observability

Observability is recommended. It is a recommended but non-mandatory key technology for moving multi-agent systems from demo to production, and should ideally cover reasoning trajectories across Skill calls, MCP tools, RAG retrieval, and LLM inference while supporting online monitoring, alerting, evaluation, and optimization.

  • If observability is used, explain data collection, semantic conventions such as OpenTelemetry GenAI, data types, backend storage and retrieval, application scenarios, and effects.
  • The observability system should cover at least one or two of Trace, Log, and Metrics.
  • Observation data should support real-time or offline evaluation and help quantify agent reasoning effectiveness and efficiency.
2.4 RAG and Context Enhancement

RAG is recommended but not mandatory. It is suitable for solutions that need to retrieve knowledge bases, history records, standards, business data, Runbooks, review evidence, or tool execution results.

  • RAG should act as a context capability within the Agent, Skill, and MCP call chain: MCP connects data sources, Skill packages retrieval, evidence alignment and result writing, and Agent judges whether retrieved result is sufficient for decision-making.
  • Based on the scenario, the solution should implement at least two of four capabilities: agent memory storage, knowledge-base RAG, shared state management, and trajectory observability.
  • If the solution explicitly does not use RAG, it should implement at least two of the remaining three capabilities excluding knowledge-base RAG and justify the effectiveness of its context mechanism.

03 Recommended Toolchain and Resource Usage

The solution should explain the open-source projects or cloud products used, their versions or compatibility scope, invocation methods, and their relationship to Agent, Skill, MCP, and RAG.

The preliminary round mainly evaluates solution design and does not require runnable code, but multi-agent collaboration design must use AgentTeams, formerly Hiclaw, as the design baseline. Teams shall explain how role orchestration, task decomposition, context passing, collaborative execution, and state tracking map to the framework capability.

Technical AreaOpen-source Project / ToolRequirementOfficial SiteDescription
Multi-agent Collaboration Framework
AgentTeams (formerly Hiclaw)
Required
hiclaw.io
Use it as the design baseline and explain how role orchestration, task decomposition, context passing, collaborative execution, and state tracking map to the framework.
Cloud Skills
Cloud Skills Portal
Recommended
skills.aliyun.com
Use it as a basic method for agents to operate cloud resources, especially authentication, orchestration, and end-to-end experience.
AI Management Center
Nacos
Recommended
nacos.io
Can be used for AI applications, agents, Skills, prompts, configuration, service discovery, or runtime governance; focus on governance design and interface relationships.
AI Gateway
Higress
Recommended
higress.io
Can be used as a unified entry point for model services, agent services, or external tool calls, including authentication, routing, rate limiting, and observability.
Agent Data Layer
PolarDB for PostgreSQL
Recommended
PolarDB-for-PostgreSQL
Can support vectors, long-term memory, RAG, audit logs, and other storage systems; explain data model, index design, permission boundaries, and replaceability.
UnifiedModel
Recommended
alibaba.github.io/UnifiedModel/
Can consistently model, associate, and query entities, data, relationships, and storage across systems.
Message Queue
RocketMQ
Recommended
rocketmq
Can support event-driven execution, async task scheduling, inter-agent messaging, execution-state transitions, and reliable notifications.
Observability
Open-source LoongSuite + AgentScope Studio or Alibaba Cloud AgentLoop
Recommended
loongsuiteagentscope-studioAlibaba Cloud AgentLoop
Can record agent reasoning trajectories and provide data support for quality, performance, cost, and reliability improvements.
  • Required: AgentTeams.
  • Recommended: official Alibaba Cloud Skills, Nacos, Higress, PolarDB for PostgreSQL, RocketMQ, LoongSuite, AgentScope Studio, and AgentLoop.
  • Alternative solutions are allowed, but teams should explain interface compatibility, replacement rationale, and migration cost.
  • Recommended projects and cloud products are not scored by quantity. Reviewers focus on design rationale, interface contracts, necessity, replaceability, permission boundaries, end-to-end loop evidence, and later migration cost.

Competition Schedule

This track includes preliminary, semi-final, and final stages. Details are as follows.

StageTimeKey Milestone / ResultKey Deliverables
Registration OpensJul 15Competition registration startsRegistration information
PreliminaryJul 15-Aug 16Preliminary submission deadline on Aug 16Project introduction and proposal deck; see preliminary submission requirements
Preliminary ReviewAug 17-Aug 24Semi-finalist list announced on Aug 24; Top 30 teams advance to the semi-finalNo additional submission unless otherwise required by the committee
Semi-finalAug 25-Sep 3Semi-final submission deadline on Sep 3Updated project proposal, executable AgentTeams code package, runnable demo / demo video; see semi-final submission requirements
Semi-final ReviewSep 4-Sep 10Finalist list announced on Sep 10; Top 15 teams advance to the final and join the on-site defenseNo additional submission unless otherwise required by the committee
FinalSep 22On-site final defense and showcaseFinal pitch deck, on-site demo, final code repository or equivalent verifiable engineering materials; see final submission requirements
GOAI DAY / AwardsSep 23Awards, project showcase, and ecosystem matchingFinal pitch deck and on-site demo

*The schedule may be adjusted flexibly according to actual progress and is subject to the committee’s final notice.

Stage Goals and Submission Materials

01 Preliminary: Direction and Solution Design

The preliminary round focuses on project direction, technical solution, open / open-source value, and feasibility. Runnable code is not required.

Submission MaterialRequiredFormatContent Description
Project IntroductionYesText, within 500 Chinese characters equivalentProject name around 20 Chinese characters equivalent; explain problem and scenario, core solution, innovation and differentiation, open / reusable value, and current progress.
Proposal DeckYesPPT / PDFRecommended content includes scenario and value, solution design, Skill and tool integration, feasibility and implementation plan, agent roles, task decomposition, context passing, result verification, exception branches, safety boundaries, risk control, and open / open-source plan.
Executable AgentTeams Code PackageOptionalRepository link / archiveNot mandatory. If submitted, include entry point, dependency instructions, configuration, sample inputs and outputs, and runtime evidence.

Semi-finalist teams will enter the next stage according to committee notices and may receive materials, Q&A, mentor support, or resources according to the competition schedule. The Organizing Committee will release further details regarding the exact number of semi-finalists, resource distribution methods and optimization suggestions via email, community groups and other channels.

02 Semi-final: Demo Implementation and Engineering Verification

The semi-final focuses on project completeness, demo runnability, engineering implementation, evaluation results, and open / open-source standards.

Submission MaterialRequiredFormatContent Description
Updated Project ProposalYesPPT / PDFUpdate the scenario loop, architecture, Skill design, risk boundary, and implementation plan based on preliminary feedback; show a complete runnable scenario chain, sample inputs and outputs, logs, Trace, metrics, evaluation results, automated verification evidence, tool / MCP / RAG / observability integration, interface schema, data flow, deployment configuration, failure handling, permissions, approval, rollback, audit mechanisms, and open / open-source plan.
Executable AgentTeams Code PackageYesRepository link / archiveInclude entry point, dependency instructions, configuration, sample inputs and outputs, and runtime evidence.
Runnable Demo / Demo VideoYesOnline experience / local deployment / video demoShow at least one complete scenario chain that verifies agent collaboration, tool calling, outputs, exception handling, execution evidence, and key technical highlights.

Finalist teams will be invited to the on-site final defense for pitches, demos, and expert Q&A. The finalist list, defense schedule, test requirements, and showcase rules will be published on the official website and notified by email or community channels.

03 Final: On-site Pitch and Project Showcase

The final focuses on presentation quality, on-site defense capability, open / open-source value, implementation potential, and long-term growth.

Judges will focus on whether the work truly completes an end-to-end task loop and has runnable, verifiable, auditable, and continuously evolvable capabilities.

Submission MaterialRequiredFormatContent Description
Final Pitch DeckYesPPT / PDFBuild a complete narrative for the on-site pitch, expert defense, and open / open-source plan, covering scenario value, solution and technical highlights, runtime results and verification evidence, engineering maturity and safety boundaries, defense support materials, and long-term plan.
On-site DemoYesOn-site demo / online experience / local deploymentShow at least one complete scenario chain that verifies agent collaboration, tool calling, outputs, and exception handling. Teams may use the Element chat room built into AgentTeams or a self-developed WebUI / other visualization.
Final Code Repository or Equivalent Verifiable Engineering MaterialsYesRepository link / accessible review packageProvide a final accessible version including README, deployment instructions, open-source license, sample configuration, and testing method.
Industry / Research Matching NeedsIf anyFormIf commercialization support is needed, describe target industry scenarios, enterprise partners, research institutions, investors, or open-source communities; otherwise this is not required.

Review Focus

1.1 General Review Dimensions

Review DimensionWeightDescription
Scenario Value and Industry Replicability25%Whether the project targets a real, clear, and representative scenario problem; clearly explains target users, pain points, real demand, and value; creates perceivable value in efficiency, cost, quality, experience, risk control, knowledge capture, or management innovation; and can be replicated, migrated, or promoted in similar industries, organizations, workflows, or user groups.
Multi-agent Collaboration and Autonomous Closed Loop25%Whether agent responsibilities are clear; task decomposition is reasonable; inter-agent context passing, collaboration protocol, and state transition are structured; exceptions, conflicts, and multiple options can be handled; and high-risk actions have human confirmation, approval, rollback, and audit boundaries.
Skill Engineering System and Ecosystem Reuse25%Whether Skills cover key task capabilities; inputs, outputs, invocation conditions, dependent tools, and failure handling are clear; Skills can be reused by multiple agents or scenarios; versioning, release, rollback, and quality evaluation are considered; and AgentTeams is used as the collaboration baseline with proper use of official Alibaba Cloud Skills.
Engineering Implementation, Runtime Verification, and Security Auditability20%Whether the demo or engineering materials are runnable; deployment, configuration, model, and tool dependencies are clear; logs, Trace, Metrics, runtime reports, and other evidence are provided; MCP / RAG / observability links have verifiable design; and data, permissions, secrets, approval, rollback, fallback, and audit mechanisms are complete.
Open-source Contribution5%Whether reusable assets, interface contracts, documentation, and examples are formed, and whether the open-source license and third-party dependencies are clear.

1.2 Track-specific Review Notes

  • AgentTeams is not evaluated by whether its name is mentioned. AgentTeams is the multi-agent collaboration design baseline, and reviewers should verify how role orchestration, task decomposition, context passing, collaborative execution, and state tracking map to the framework capability.
  • Skill is mandatory for this track. Reviewers will focus on Skill inputs and outputs, invocation conditions, dependent tools, failure handling, validation method, reuse value, version evolution, and open-source distribution design.
  • Recommended projects and cloud products are not scored by quantity. Reviewers focus on necessity, interface contract, replaceability, permission boundaries, end-to-end loop evidence, and later migration cost.

Participant Support

This track plans to provide teams that complete preliminary submission and pass submission validity review with up to RMB 200 in compute, cloud services, and other development resource subsidies, or equivalent competition support resources, with a quota of up to 300 slots in principle. After the preliminary submission deadline, the organizing committee will, based on submission and review progress, notify eligible teams via email or official community channels to submit relevant materials, and distribute resources in an orderly manner according to the rules. Resource support is not a competition prize, review bonus, or advancement criterion. Specific support forms, application conditions, material requirements, distribution methods, and schedule are subject to subsequent notices from the organizing committee.

Best-fit Teams

This track is suitable for the following teams:

  • Developer teams with multi-agent system design capabilities
  • Technical teams familiar with cloud products, enterprise systems, DevOps, ITSM, FinOps, or security-system integration
  • Teams with understanding of enterprise-grade complex scenarios such as operations, security, cost governance, and capacity governance
  • Teams familiar with business systems, data platforms, development toolchains, or industry application integration
  • Teams that understand complex task collaboration, process automation, knowledge capture, risk control, efficiency improvement, or experience optimization in real industry scenarios
  • University labs, open-source teams, enterprise technical teams, or independent AI builders with engineering implementation, toolchain integration, and demo-building capabilities
  • Teams with open-source project building, technical documentation, or community collaboration experience are preferred

Registration & Resources