7. Multi-Agent & A2A

As individual agents mature, the frontier shifts to agent collaboration — how agents communicate, coordinate, and form societies. 2025 saw major standardization efforts with Google's A2A Protocol and the AG-UI Protocol for agent-user interaction.

7.1 Evolution of Multi-Agent Systems

7.2 A2A Protocol (Google, 2025.4)

The Agent-to-Agent (A2A) Protocol is Google's open standard for inter-agent communication, announced in April 2025 with support from 50+ companies.

Why A2A Matters

Problem Without A2A	Solution With A2A
Custom integration per agent pair	Standard protocol for all
Vendor lock-in	Open, interoperable
No discovery mechanism	Agent Cards for capability discovery
Proprietary message formats	Standard JSON-RPC messages
No task lifecycle management	Standardized task states

Core Concepts

Concept	Description
Agent Card	JSON metadata describing an agent's capabilities, endpoints, and authentication
Task	A unit of work sent from one agent to another with lifecycle management
Message	Communication within a task (text, files, forms)
Part	Individual content pieces (text, file, data) within a message
Push Notification	Async notification for long-running tasks

Agent Card Example

{
  "name": "Research Agent",
  "description": "Performs web research and synthesizes findings",
  "url": "https://research-agent.example.com/a2a",
  "capabilities": {
    "streaming": true,
    "pushNotifications": true
  },
  "skills": [
    {
      "id": "web-research",
      "name": "Web Research",
      "description": "Search and synthesize information from the web"
    },
    {
      "id": "data-analysis",
      "name": "Data Analysis",
      "description": "Analyze datasets and generate reports"
    }
  ],
  "authentication": {
    "schemes": ["bearer"]
  }
}

Task Lifecycle

A2A vs MCP

Aspect	MCP (Model Context Protocol)	A2A (Agent-to-Agent)
Purpose	Agent-to-Tool communication	Agent-to-Agent communication
Analogy	USB for connecting peripherals	HTTP for connecting services
Scope	Single agent's tool ecosystem	Multi-agent collaboration
Transport	stdio / SSE	HTTP / JSON-RPC
Discovery	Server-configured	Agent Cards
Created by	Anthropic (2024)	Google (2025)
Relationship	Complementary	Complementary

7.3 AG-UI Protocol (CopilotKit, 2025)

The Agent-User Interaction Protocol (AG-UI) standardizes how agents interact with human users through UI components.

Why AG-UI Matters

Traditional agents return text. AG-UI enables agents to:

Render rich UI components (forms, tables, charts)
Stream real-time progress indicators
Request structured user input
Display interactive elements

Architecture

Event Types

Event Type	Direction	Description
TextMessageStart/Content/End	Agent → UI	Streaming text output
StateSnapshot / StateDelta	Agent → UI	Agent state updates
ToolCallStart/Args/End	Agent → UI	Tool execution progress
ToolCallResult	Agent → UI	Tool results
RunStarted / RunFinished	Agent → UI	Lifecycle events

7.4 Real-World Case Study: Docker Agent Fleet (2026)

Docker 团队在其 Coding Agent Sandboxes（sbx）项目中构建了一个由 7 个 AI Agent 角色组成的 "虚拟团队"，用于自动化 CI/CD 流程。这是多 Agent 协作在工程生产中的一个典型落地案例。

核心设计理念

角色而非脚本：每个 Agent 是一个 SKILL.md 文件，定义角色（persona）、职责和可用工具，而非固定的执行步骤
本地优先，CI 其次：所有 Skill 先在开发者本地验证，确认行为正确后再接入 CI
同一套 Skill，两个运行时：本地终端和 CI 环境运行完全相同的 Skill 文件

七个 Agent 角色

角色	职责
`/build-engineer`	构建二进制文件、容器模板、平台特定构建标志
`/project-manager`	去重发现结果、管理 GitHub Projects、自动化 Triage
`/product-owner`	将 commit 翻译为人类可读的发布说明
`/cli-tester`	探索性测试，52+ 测试场景，14 个测试层级
`/performance-tester`	生命周期耐久性测试、I/O 性能基准
`/upgrade-tester`	四阶段升级回归测试
`/software-engineer`	响应 GitHub issue 标签，自动修复 Bug

关键创新

PR 自动审查：当有人在 PR 中评论 /cli-tester-review，CI 在 MacOS、Linux、Windows 三个平台上并行运行探索性测试，结果直接作为 PR 评论发布
去重机制：project-manager 在创建 issue 前会通过 GraphQL 分页扫描整个项目板，避免重复
渐进式自主性：本地运行时交互式提问，CI 中完全自动模式

来源：Docker Blog - A Virtual Agent team at Docker（2026-05-01）

7.5 Multi-Agent Patterns

Common Orchestration Patterns

1. Supervisor Pattern

A central supervisor agent delegates tasks to specialized workers.

2. Hierarchical Pattern

Multi-level management with strategic and tactical planning.

3. Mesh / Peer-to-Peer Pattern

Agents communicate directly without a central coordinator.

4. Debate Pattern

Multiple agents discuss and reach consensus through structured debate.

Pattern Comparison

Pattern	Complexity	Scalability	Best For
Supervisor	Low	Medium	Task delegation
Hierarchical	Medium	High	Large organizations
Mesh	High	Medium	Creative collaboration
Debate	Medium	Low	Quality improvement

7.5 Agent Society & Swarms

Self-Organizing Agents

Agents that dynamically form teams based on task requirements.

Key Concepts

Concept	Description
Agent Registry	Directory of available agents and their capabilities
Dynamic Team Formation	Agents self-organize based on task requirements
Task Market	Tasks are posted and agents bid to complete them
Reputation System	Track agent performance and reliability
Economic Model	Token-based compensation for agent services

When More Is Less：多 Agent 协作何时失效（2026-09）

arXiv 2609.19759（Rethinking Multi-Agent Collaboration: When More Is Less）系统论证：随着单 Agent 能力（含 harness 质量）持续增强，多 Agent 协作并非总是带来真实增益。论文给出判别"何时多 Agent 有价值"的框架，并指出许多流行多 Agent 模式在强单体基线下收益为负。这与工程社区的实践观察相互印证——OpenAI Codex 工程师 Eric Provencher 曾警告"超过两个并行子 Agent 几乎总是在烧 token 而不提升质量"。

选型启示：在引入多 Agent 拓扑前，先跑强单 Agent + 良好 harness 的基线；多 Agent 的合理场景是任务天然可并行分解、子任务间信息依赖弱、以及需要独立视角交叉验证（如对抗性评审）时。并行度本身不是目标。

来源：arXiv:2609.19759（2026-09-17）

Agensh：把组织智能扩展到 1024 个 Agent（2026-09-22）

arXiv 2609.26781（Agensh: Scaling Organizational Intelligence to 1,024 Agents）沿"Agent 社会"路线把规模推到四位数：通过分层组织结构让上千个 Agent 协同工作而不失控。这与"When More Is Less"的警告形成有趣对照——规模本身既不是问题也不是答案，组织结构才是：无结构的堆叠 Agent 数量会放大失败模式，而合适的分层与角色设计可以在大规模下保持增益。

与同期 2609.26780（SpeakerMem-R1）合看还有一层启示：多方（多 Agent）协作的记忆瓶颈与人类群聊一致——谁说了什么、对谁说的、群体共享了什么。SpeakerMem 的"说话人标签逐字消息 + 个人/群组双视图"设计对多 Agent 系统的共享记忆同样适用。

来源：arXiv:2609.26781、arXiv:2609.26780（2026-09-22）

7.6 W3C ANP (Agent Network Protocol)

The W3C is working on standardizing agent networking through the Agent Network Protocol (ANP).

Goals

Interoperability: Agents from different vendors can communicate
Discovery: Standard mechanism for finding agents
Trust: Verification and reputation systems
Privacy: Data sharing controls
Security: Authentication and authorization

Relationship to Other Standards

W3C ANP (Agent Network Protocol)
  ├── Builds on: A2A Protocol (Google)
  ├── Complements: MCP (Anthropic)
  ├── Inspired by: AG-UI (CopilotKit)
  └── Aligns with: Web Standards (HTTP, JSON-LD)

7.7 Implementation with Agent SDKs

OpenAI Agents SDK — Multi-Agent Handoff

from agents import Agent, Runner

researcher = Agent(
    name="Researcher",
    instructions="You research topics thoroughly.",
)

writer = Agent(
    name="Writer",
    instructions="You write clear, engaging content.",
)

coordinator = Agent(
    name="Coordinator",
    instructions="Route tasks to the right specialist.",
    handoffs=[researcher, writer],
)

result = Runner.run_sync(coordinator, "Write a report on quantum computing")

LangGraph — Supervisor Pattern

from langgraph.graph import StateGraph, END
from typing import TypedDict, Annotated
import operator

class State(TypedDict):
    messages: Annotated[list, operator.add]
    next: str

def supervisor(state: State):
    # Decide next agent
    return {"next": "researcher"}

def researcher(state: State):
    # Research and return findings
    return {"messages": ["Research findings..."]}

workflow = StateGraph(State)
workflow.add_node("supervisor", supervisor)
workflow.add_node("researcher", researcher)
workflow.add_conditional_edges("supervisor", lambda s: s["next"])
workflow.add_edge("researcher", "supervisor")
workflow.set_entry_point("supervisor")

app = workflow.compile()

7.8 Multi-Agent 安全与基础设施

LCGuard：多 Agent 潜在通信安全（2026）

论文 LCGuard 关注 LLM 多 Agent 系统中一个新兴的安全问题：当 Agent 之间通过 KV Cache（而非自然语言）共享中间状态时，可能产生潜在信道泄露。

核心问题：

多 Agent 系统中，为提高效率开始使用共享 KV Cache 进行通信
KV Cache 中可能编码了敏感信息（用户隐私、内部推理过程）
恶意 Agent 可以通过分析 KV Cache 的统计特征推断其他 Agent 的输入

LCGuard 方案：通过在 KV 共享层添加潜在的通信守卫，过滤敏感信息的同时保持协作效率。

来源：arXiv:2605.22786（2026-05）

DeltaBox：毫秒级沙盒检查点/回滚（2026）

论文 DeltaBox 解决了 LLM Agent 高频状态探索的基础设施问题——为 AI Agent 提供毫秒级的完整沙盒状态检查点和回滚能力。

应用场景：

测试时树搜索（test-time tree search）
强化学习训练中的 episode 回滚
Agent 错误恢复和状态探索

技术亮点：基于增量的 C/R 机制，避免了传统虚拟机快照的巨大开销，使 Agent 可以在毫秒级保存和恢复完整的文件系统 + 进程状态。

来源：arXiv:2605.22781（2026-05）

7.9 Key Takeaways

A2A Protocol is becoming the standard for inter-agent communication
AG-UI Protocol bridges the gap between agents and rich user interfaces
MCP + A2A are complementary — tools vs. agent collaboration
Multi-agent patterns (Supervisor, Hierarchical, Mesh) solve different coordination needs
Agent societies represent the frontier — self-organizing, economic systems

Start with Patterns

Before building multi-agent systems, master the Supervisor pattern — it's the most practical for most use cases. Use LangGraph or OpenAI Agents SDK for implementation.

Protocol Selection

Need agent-to-tool communication? Use MCP
Need agent-to-agent communication? Use A2A
Need agent-to-user rich interaction? Use AG-UI

7.1 Evolution of Multi-Agent Systems​

7.2 A2A Protocol (Google, 2025.4)​

Why A2A Matters​

Core Concepts​

Agent Card Example​

Task Lifecycle​

A2A vs MCP​

7.3 AG-UI Protocol (CopilotKit, 2025)​

Why AG-UI Matters​

Architecture​

Event Types​

7.4 Real-World Case Study: Docker Agent Fleet (2026)​

核心设计理念​

七个 Agent 角色​

关键创新​

7.5 Multi-Agent Patterns​

Common Orchestration Patterns​

1. Supervisor Pattern​

2. Hierarchical Pattern​

3. Mesh / Peer-to-Peer Pattern​

4. Debate Pattern​

Pattern Comparison​

7.5 Agent Society & Swarms​

Self-Organizing Agents​

Key Concepts​

When More Is Less：多 Agent 协作何时失效（2026-09）​

Agensh：把组织智能扩展到 1024 个 Agent（2026-09-22）​

7.6 W3C ANP (Agent Network Protocol)​

Goals​

Relationship to Other Standards​

7.7 Implementation with Agent SDKs​

OpenAI Agents SDK — Multi-Agent Handoff​

LangGraph — Supervisor Pattern​

7.8 Multi-Agent 安全与基础设施​

LCGuard：多 Agent 潜在通信安全（2026）​

DeltaBox：毫秒级沙盒检查点/回滚（2026）​

7.9 Key Takeaways​