BrainBank
AI Classroom/MCPClaude Code Deep Dive

SendMessageTool: Agent Communication

8/2/2026, 4:37:57 PM · updated 8/8/2026, 10:44:59 AM · Source

AI-translated on 8/2/2026, 4:40:05 PM · by Qwen3.6 35B (fast, default)

#mcp#claude-code#multi-agent#send-message-tool#structured-message#agent-communication

The retrieved knowledge base entries do not contain specific information about **SendMessageTool** within Claude Code's multi-agent architecture, including its message protocol design, input schema, or communication mechanisms. The results focus on general computing concepts such as network protocols (e.g., packet transmission in CSN-1.C), student response systems, and biological signaling—none of which directly address the technical details requested. Given that your recent activity includes work on **K12/Agent toolchain task suspension stop mechanism** and related Claude Code components (*detailed under "Recently working on"* above), this specific gap suggests either: 1) The topic may not yet be documented in the current knowledge base, or 2) A more targeted search (e.g., using a narrower query like "Claude Code SendMessageTool input schema" or expanding to external tools) would be necessary to locate relevant details. I cannot fabricate or extrapolate technical specifications from unrelated materials. If this is blocking your progress, consider: - Using `mcp__tavily__tavily_search` for real-time web research on Claude Code documentation. - Reviewing internal skill files like `K12/GlobTool design principles FileEditTool FileWriteTool` (if they contain toolchain context). Would you like me to attempt a more refined search or explore external tools?

It Serves as the Communication Bus in Multi-Agent Mode

In Claude Code's multi-agent architecture, plain text is not automatically visible to other teammates.
So if an agent wants to notify another agent, broadcast a message, or respond to an approval, it must route through SendMessageTool.

This means the tool isn't an auxiliary feature, but the foundational infrastructure that makes a true multi-agent system possible.

Key Source Code

tools/SendMessageTool/SendMessageTool.ts:

const inputSchema = z.object({
  to: z.string(),
  summary: z.string().optional(),
  message: z.union([z.string(), StructuredMessage()]),
})

And StructuredMessage supports several types of special messages:

type: 'shutdown_request'
type: 'shutdown_response'
type: 'plan_approval_response'

This indicates that it doesn't just send plain text, but already supports protocol-formatted messaging.

Message Routing Chain

An agent needs to contact someone > SendMessageTool > Resolve to address > Write to mailbox / bridge / peer > Target agent receives the message > Continue collaboration or update state

It Supports More Than Just Teammate Names

Learning map

Learning Roadmap

Phase One: Conceptual Foundation — Why Multi-Agent Systems Need Dedicated Communication?

  • Understand that agents cannot share context text by default (each conversation is isolated)
  • Distinguish the scope differences between "plain text" and "protocol messages"
  • Understand the basic architecture of the Claude Code multi-agent system (Lead Agent → Teammates)

Phase Two: Core Source Code — Message Schema of SendMessageTool

  • Master the meaning and constraints of the input parameters to, summary, and message
  • Understand the StructuredMessage type hierarchy (shutdown_request/response, plan_approval_response, etc.)
  • Grasp how the tool upgrades messages from "free-form text" to "protocol messages"

Phase Three: Invocation Path — How a Message Reaches Its Target Agent

  • SendMessageTool → routing/parsing of the to address → mailbox / bridge / peer → delivery to target agent
  • Understand how the request-response model drives collaboration and approvals in agent negotiations

Phase Four: Advanced Scenarios — Applications Beyond Teammate Names

  • Cross-machine multi-agent collaboration (Shared Channel + Real-time Messaging)
  • Enhance agent-to-agent protocol security via MCP (authentication, approval workflows)
  • Integrate with companion tools such as TaskCreate/UpdateTool and TeamCreate/DeleteTool to implement complete team workflows

Get hands-on — step by step

Hands-on Exercise

  1. Install and start Claude Code (run claude in the terminal), ensuring the version includes multi-Agent support.
  2. Open a real project directory, activate Claude Code in the terminal, and enter a task request that triggers multi-Agent behavior (e.g., "Help me split this PR's code review and documentation tasks, assigning them to two sub-agents for parallel processing").
  3. Observe Claude Code's Agent allocation output—confirm that the Lead Agent created Teammate A and Teammate B.
  4. In the Lead Agent's response, find the SendMessageTool tool call record (the tool block starting with sendMessage or Send Message), and parse its parameters:
    • to: the target Agent's name / ID
    • message: the command or structured message fields it carries
  5. Switch terminal windows or continue in the same session to trigger an approval scenario request (e.g., "Please modify the API routes under src/; I'll review them after you're done")—observe how the Lead Agent sends a plan_approval_response type message via SendMessageTool to the dispatched sub-agent.
  6. (Advanced) Clone the willin/claude-code-source-code repository, locate tools/SendMessageTool/SendMessageTool.ts:
    • Read the complete input Schema definition
    • Trace the communication logic of mailbox and worker in the bridge/ and coordinator/ directories
  7. (Advanced) Attempt to check the cross-machine communication design plan at Anthropic Claude Code GitHub issues #28300 (Agent-to-Agent protocol), and understand how Shared Channel + Real-time Messaging expands the capabilities boundary of SendMessageTool.

Top 3 sources

  1. 1
    willin/claude-code-source-code — Claude Code v2.1.88 源码仓库

    逐文件整理并文档化的 Claude Code 源码库，README 中明确列出了 SendMessageTool 在通信总线架构中的位置、调用链与多 Agent 协作机制。

    https://github.com/willin/claude-code-source-code

  2. 2
    Anthropic claude-code Issue #28300 — Agent-to-Agent Protocol

    Anthropic 官方 issue，阐述多机器跨 Agent 通信协议设计——Shared Channel、Real-time Messaging 和安全审批机制的演进方向。

    https://github.com/anthropics/claude-code/issues/28300

  3. 3
    Claude Code 官方文档

    Anthropic 官方发布的 Claude Code 使用指南，涵盖 Agent 工具调用、多智能体协作模式与 MCP Server 集成的完整说明。

    https://docs.anthropic.com/en/docs/claude-code/overview

Links are AI-suggested — worth a quick sanity check before diving in.