Add multi-agent coordination diagrams

Author: syhya

content/en/posts/2026-06-21-agent-architecture-design/index.md

@@ -260,12 +260,26 @@ The effectiveness of a Multi-Agent System depends heavily on the lead agent's ta
 
 When a workflow is driven by dynamic events and conditional logic accumulates in the central coordinator, a **Message Bus** ([Phillips, 2026](https://claude.com/blog/multi-agent-coordination-patterns)) makes routing explicit and extensible. Agents publish events and subscribe to relevant topics, while a router delivers matching messages; new Agent types can join existing topics without rewiring connections among the existing Agents.
 
+{{< figure
+    src="message-bus.png"
+    caption="Fig. 16. Message Bus pattern: Agents publish events to a shared bus and subscribe to relevant topics, remaining decoupled from one another. (Image source: [Phillips, 2026](https://claude.com/blog/multi-agent-coordination-patterns))"
+    align="center"
+    width="100%"
+>}}
+
 Message Bus works well for security alerts, asynchronous tasks, and event-driven systems whose Agent ecosystem is likely to grow. The trade-off is harder tracing: if the router misclassifies or drops an event, or a downstream subscription is misconfigured, work may be omitted silently without a crash. Event chains therefore need complete, correlated logs, and the results of LLM-based routing should be validated.
 
 ### Shared State
 
 When Agents need to build on one another's intermediate findings rather than respond to discrete events, they can use **Shared State** ([Phillips, 2026](https://claude.com/blog/multi-agent-coordination-patterns)). All Agents directly read and write the same database, file system, or document. With no Orchestrator or router deciding where each finding goes, the shared store itself becomes an evolving collaboration context.
 
+{{< figure
+    src="shared-state.png"
+    caption="Fig. 17. Shared State pattern: autonomous Agents read from and write to a common persistent store, making findings immediately available to the other Agents. (Image source: [Phillips, 2026](https://claude.com/blog/multi-agent-coordination-patterns))"
+    align="center"
+    width="100%"
+>}}
+
 Shared State works well for collaborative research and shared knowledge bases, but removing the central coordinator can also lead to duplicate work, conflicting writes, and reactive loops. Locking, versioning, or partitioning can mitigate concurrent writes; behavioral loops require explicit termination conditions, such as a time budget, several rounds without new findings, or a designated Agent deciding that the available information is sufficient. In short, use Message Bus when discrete events trigger downstream actions, and Shared State when findings must accumulate and be revisited.
 
 ### Coordination Boundaries
@@ -276,7 +290,7 @@ In a software-vulnerability experiment, Anthropic launched **45 Agents**, each i
 
 {{< figure
     src="multi-agent-vulnerability-discovery.png"
-    caption="Fig. 16. Coordinated swarm findings (solid curves) versus independently partitioned agents (stars); dashed curves show overlap, while the dotted Mythos Preview curve keeps only findings inside the baseline's core directories. Search scopes and token budgets are not matched. (Image source: [Anthropic, 2026e](https://www.anthropic.com/research/multiagent-systems))"
+    caption="Fig. 18. Coordinated swarm findings (solid curves) versus independently partitioned agents (stars); dashed curves show overlap, while the dotted Mythos Preview curve keeps only findings inside the baseline's core directories. Search scopes and token budgets are not matched. (Image source: [Anthropic, 2026e](https://www.anthropic.com/research/multiagent-systems))"
     align="center"
     width="100%"
 >}}
@@ -287,7 +301,7 @@ Shared code presents a different challenge because each edit immediately changes
 
 {{< figure
     src="multi-agent-code-coordination.png"
-    caption="Fig. 17. End-of-run merged-PR fraction and median-agent code sharing for 12-hour game-building swarms, averaged across three prompt types. These are coordination proxies rather than quality scores, and high merge rates can reflect file isolation instead of collaboration. (Image source: [Anthropic, 2026e](https://www.anthropic.com/research/multiagent-systems))"
+    caption="Fig. 19. End-of-run merged-PR fraction and median-agent code sharing for 12-hour game-building swarms, averaged across three prompt types. These are coordination proxies rather than quality scores, and high merge rates can reflect file isolation instead of collaboration. (Image source: [Anthropic, 2026e](https://www.anthropic.com/research/multiagent-systems))"
     align="center"
     width="100%"
 >}}
@@ -322,7 +336,7 @@ Viewing it alongside Subagents and Agent Teams, the three represent different or
 
 {{< figure
     src="compare-agent-teams-dynamic-workflows.png"
-    caption="Fig. 18. Claude Code parallelizes work in several ways: subagents delegate a side task in their own context and return only a summary, agent teams coordinate through a shared task list with direct messaging, and dynamic workflows script and cross-check many subagents for jobs too big to coordinate one turn at a time. (Image source: [Cat Wu on X](https://x.com/_catwu/status/2060054180379689074))"
+    caption="Fig. 20. Claude Code parallelizes work in several ways: subagents delegate a side task in their own context and return only a summary, agent teams coordinate through a shared task list with direct messaging, and dynamic workflows script and cross-check many subagents for jobs too big to coordinate one turn at a time. (Image source: [Cat Wu on X](https://x.com/_catwu/status/2060054180379689074))"
     align="center"
     width="100%"
 >}}
@@ -333,7 +347,7 @@ Anthropic summarizes several common workflow patterns. When Claude generates a h
 
 {{< figure
     src="six-workflow-patterns.png"
-    caption="Fig. 19. Six common dynamic workflow patterns in Claude Code: classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, and loop until done. (Image source: [Anthropic, 2026d](https://claude.com/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code))"
+    caption="Fig. 21. Six common dynamic workflow patterns in Claude Code: classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, and loop until done. (Image source: [Anthropic, 2026d](https://claude.com/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code))"
     align="center"
     width="100%"
 >}}
@@ -355,7 +369,7 @@ The preceding sections introduced increasingly complex modes of organization, bu
 
 {{< figure
     src="agent-archs.png"
-    caption="Fig. 20. The agent system architectures studied: a single-agent baseline alongside several multi-agent coordination topologies. (Image source: [Kim et al., 2025](https://arxiv.org/abs/2512.08296))"
+    caption="Fig. 22. The agent system architectures studied: a single-agent baseline alongside several multi-agent coordination topologies. (Image source: [Kim et al., 2025](https://arxiv.org/abs/2512.08296))"
     align="center"
     width="100%"
 >}}
@@ -365,7 +379,7 @@ Several key findings:
 
 {{< figure
     src="agent-scaling.png"
-    caption="Fig. 21. Multi-agent coordination helps on parallelizable tasks but hurts on sequential ones, and the benefit shrinks as the single-agent baseline grows stronger (capability saturation). (Image source: [Kim et al., 2025](https://arxiv.org/abs/2512.08296))"
+    caption="Fig. 23. Multi-agent coordination helps on parallelizable tasks but hurts on sequential ones, and the benefit shrinks as the single-agent baseline grows stronger (capability saturation). (Image source: [Kim et al., 2025](https://arxiv.org/abs/2512.08296))"
     align="center"
     width="100%"
 >}}

content/en/posts/2026-06-21-agent-architecture-design/message-bus.png

@@ -260,12 +260,26 @@ Multi-Agent Systems 的效果很大程度取决于 lead agent 的任务拆解能
 
 当工作流由动态事件驱动、中心协调者的条件分支越来越复杂时，可以引入 **Message Bus（消息总线）**（[Phillips, 2026](https://claude.com/blog/multi-agent-coordination-patterns)）：Agent 只负责 publish 事件或 subscribe 自己关心的 topic，由 router 完成消息分发，新 Agent 可以接入已有 topic，而不需要重写其他 Agent 的连接关系。
 
+{{< figure
+    src="message-bus.png"
+    caption="Fig. 16. Message Bus pattern: Agents publish events to a shared bus and subscribe to relevant topics, remaining decoupled from one another. (Image source: [Phillips, 2026](https://claude.com/blog/multi-agent-coordination-patterns))"
+    align="center"
+    width="100%"
+>}}
+
 Message Bus 适合安全告警、异步任务和不断增加新 Agent 类型的事件驱动系统。它的代价是执行链路更难追踪：router 分类错误、漏投事件或下游订阅配置错误时，系统可能没有显式崩溃，却静默地遗漏工作。因此需要为事件链保留完整、可关联的日志，并验证 LLM router 的路由结果。
 
 ### Shared State
 
 如果 Agent 不是依次响应事件，而是需要持续利用彼此的中间发现，可以改用 **Shared State（共享状态）**（[Phillips, 2026](https://claude.com/blog/multi-agent-coordination-patterns)）。所有 Agent 直接读写同一个数据库、文件系统或文档，不再由 orchestrator 或 router 决定信息应该传给谁；共享存储本身成为持续演化的协作上下文。
 
+{{< figure
+    src="shared-state.png"
+    caption="Fig. 17. Shared State pattern: autonomous Agents read from and write to a common persistent store, making findings immediately available to the other Agents. (Image source: [Phillips, 2026](https://claude.com/blog/multi-agent-coordination-patterns))"
+    align="center"
+    width="100%"
+>}}
+
 Shared State 适合协作研究和共同维护知识库，但去掉中心协调者也会带来重复工作、矛盾写入和 reactive loop。并发写入可以用锁、版本控制或分区缓解；行为层面的循环则必须依靠明确的停止条件，例如时间预算、连续若干轮没有新发现，或由指定 Agent 判断信息已经充分。简而言之，离散事件触发后续动作时使用 Message Bus；信息需要持续积累并被反复读取时使用 Shared State。
 
 ### 协调边界
@@ -276,7 +290,7 @@ Shared State 适合协作研究和共同维护知识库，但去掉中心协调
 
 {{< figure
     src="multi-agent-vulnerability-discovery.png"
-    caption="Fig. 16. Coordinated swarm findings (solid curves) versus independently partitioned agents (stars); dashed curves show overlap, while the dotted Mythos Preview curve keeps only findings inside the baseline's core directories. Search scopes and token budgets are not matched. (Image source: [Anthropic, 2026e](https://www.anthropic.com/research/multiagent-systems))"
+    caption="Fig. 18. Coordinated swarm findings (solid curves) versus independently partitioned agents (stars); dashed curves show overlap, while the dotted Mythos Preview curve keeps only findings inside the baseline's core directories. Search scopes and token budgets are not matched. (Image source: [Anthropic, 2026e](https://www.anthropic.com/research/multiagent-systems))"
     align="center"
     width="100%"
 >}}
@@ -287,7 +301,7 @@ Agent swarm 使用了约四倍的 Token，并扩展到了更多代码目录；
 
 {{< figure
     src="multi-agent-code-coordination.png"
-    caption="Fig. 17. End-of-run merged-PR fraction and median-agent code sharing for 12-hour game-building swarms, averaged across three prompt types. These are coordination proxies rather than quality scores, and high merge rates can reflect file isolation instead of collaboration. (Image source: [Anthropic, 2026e](https://www.anthropic.com/research/multiagent-systems))"
+    caption="Fig. 19. End-of-run merged-PR fraction and median-agent code sharing for 12-hour game-building swarms, averaged across three prompt types. These are coordination proxies rather than quality scores, and high merge rates can reflect file isolation instead of collaboration. (Image source: [Anthropic, 2026e](https://www.anthropic.com/research/multiagent-systems))"
     align="center"
     width="100%"
 >}}
@@ -322,7 +336,7 @@ Dynamic Workflows 要解决的，正是长程任务在单一上下文中反复
 
 {{< figure
     src="compare-agent-teams-dynamic-workflows.png"
-    caption="Fig. 18. Claude Code parallelizes work in several ways: subagents delegate a side task in their own context and return only a summary, agent teams coordinate through a shared task list with direct messaging, and dynamic workflows script and cross-check many subagents for jobs too big to coordinate one turn at a time. (Image source: [Cat Wu on X](https://x.com/_catwu/status/2060054180379689074))"
+    caption="Fig. 20. Claude Code parallelizes work in several ways: subagents delegate a side task in their own context and return only a summary, agent teams coordinate through a shared task list with direct messaging, and dynamic workflows script and cross-check many subagents for jobs too big to coordinate one turn at a time. (Image source: [Cat Wu on X](https://x.com/_catwu/status/2060054180379689074))"
     align="center"
     width="100%"
 >}}
@@ -333,7 +347,7 @@ Anthropic 总结了几种常见的 workflow pattern。Claude 在运行时生成
 
 {{< figure
     src="six-workflow-patterns.png"
-    caption="Fig. 19. Six common dynamic workflow patterns in Claude Code: classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, and loop until done. (Image source: [Anthropic, 2026d](https://claude.com/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code))"
+    caption="Fig. 21. Six common dynamic workflow patterns in Claude Code: classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, and loop until done. (Image source: [Anthropic, 2026d](https://claude.com/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code))"
     align="center"
     width="100%"
 >}}
@@ -355,7 +369,7 @@ Anthropic 总结了几种常见的 workflow pattern。Claude 在运行时生成
 
 {{< figure
     src="agent-archs.png"
-    caption="Fig. 20. The agent system architectures studied: a single-agent baseline alongside several multi-agent coordination topologies. (Image source: [Kim et al., 2025](https://arxiv.org/abs/2512.08296))"
+    caption="Fig. 22. The agent system architectures studied: a single-agent baseline alongside several multi-agent coordination topologies. (Image source: [Kim et al., 2025](https://arxiv.org/abs/2512.08296))"
     align="center"
     width="100%"
 >}}
@@ -365,7 +379,7 @@ Anthropic 总结了几种常见的 workflow pattern。Claude 在运行时生成
 
 {{< figure
     src="agent-scaling.png"
-    caption="Fig. 21. Multi-agent coordination helps on parallelizable tasks but hurts on sequential ones, and the benefit shrinks as the single-agent baseline grows stronger (capability saturation). (Image source: [Kim et al., 2025](https://arxiv.org/abs/2512.08296))"
+    caption="Fig. 23. Multi-agent coordination helps on parallelizable tasks but hurts on sequential ones, and the benefit shrinks as the single-agent baseline grows stronger (capability saturation). (Image source: [Kim et al., 2025](https://arxiv.org/abs/2512.08296))"
     align="center"
     width="100%"
 >}}