SOP Agents

Multi-agent orchestration for Strands Agents, defined in plain text. Write what your agents do. Write how they collaborate. Done.

📖 Read the announcement: Introducing Strands Agent SOPs on the AWS Open Source Blog

<!-- sops/researcher.md -->
---
name: researcher
description: Researches topics and gathers information
version: 1.0.0
type: agent
---
# Research Agent

## Overview

You are a research specialist focused on gathering accurate information.

## Steps

### 1. Analyze the Topic

Understand what information is needed and identify reliable sources.

**Constraints:**
- You MUST use authoritative sources
- You SHOULD cross-reference multiple sources

### 2. Gather and Synthesize

Collect relevant information and organize it clearly.

**Constraints:**
- You MUST cite your sources
- You MUST distinguish facts from speculation

<!-- sops/writer.md -->
---
name: writer  
description: Writes content based on research and requirements
version: 1.0.0
type: agent
---
# Writer Agent

## Overview

You are a skilled writer who transforms research into engaging content.

## Steps

### 1. Review the Input

Understand the research findings and target audience.

**Constraints:**
- You MUST NOT invent facts because the output must be grounded in provided research
- You SHOULD adapt tone to the target audience

### 2. Write and Polish

Create clear, engaging content from the source material.

**Constraints:**
- You MUST maintain factual accuracy
- You SHOULD use clear, accessible language

// That's it. Now they work together automatically.
const orchestrator = await createOrchestrator({ directory: "./sops" });

const result = await orchestrator.invoke(
  "Research quantum computing breakthroughs in 2025 and write a blog post about them"
);
// The orchestrator reads your request, calls researcher, passes results to writer,
// and returns the finished blog post.

No glue code. No tool definitions. No agent-to-agent wiring. Just describe what each agent does, and the orchestrator figures out how to coordinate them.

Why?

Building multi-agent systems usually means:

• Writing tool schemas and type definitions
• Manually wiring agents together
• Hardcoding orchestration logic
• Rebuilding when requirements change

With SOP Agents, you write markdown files. The library handles the rest.

Model-driven orchestration means the LLM decides how to coordinate agents at runtime—not you at build time. Traditional approaches lock you into predefined flows. Here, the orchestrator reasons about each request and picks the best path. Add a new agent? It's immediately available. Rephrase your request? The orchestrator adapts. No rewiring required.

Install

npm install @serverless-dna/sop-agents

Requires AWS credentials for Amazon Bedrock (the default model provider):

aws login  # or aws configure

How It Works

1. You write SOPs - Markdown files describing what each agent does
2. Library generates tools - Each SOP becomes a callable tool with typed inputs
3. Orchestrator coordinates - A master agent reads requests and delegates to the right agents
4. Agents execute - Each agent runs with its SOP as the system prompt

The orchestrator is itself an agent—defined the same way. It sees your request, looks at the available agent tools, and decides which to call and in what order. No flowcharts, no state machines—just LLM reasoning.

Out of the box, the library includes a default orchestrator that analyzes your request and selects the right agent(s) to handle it. One agent, multiple agents in sequence, or a back-and-forth—it figures out what's needed.

Want more control? Write your own orchestrator SOP:

<!-- sops/orchestrator.md -->
---
name: orchestrator
description: Coordinates research and writing tasks
type: orchestrator
---
# Content Pipeline Orchestrator

You coordinate a content creation pipeline.

## Steps

1. Always start with the researcher to gather facts
2. Pass research findings to the writer
3. If the writer needs clarification, go back to the researcher
4. Review the final output for accuracy before returning

Never skip research. Never let the writer make up facts.

Plain English orchestration. Override the default when you need specific workflows.

Using MCP Tools

Agents can access external tools and services through the Model Context Protocol (MCP). MCP provides a standardized way to connect agents to databases, APIs, file systems, and other capabilities.

Registering Tools

Pass MCP clients as a named registry when creating the orchestrator. Each agent then declares which tools it needs in its frontmatter:

import { createOrchestrator } from "@serverless-dna/sop-agents";
import { McpClient } from "@strands-agents/sdk";
import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js";

async function main() {
  // Create MCP clients
  const fileSystem = new McpClient({
    transport: new StdioClientTransport({
      command: "npx",
      args: ["-y", "@modelcontextprotocol/server-filesystem", "./workspace"],
    }),
  });

  const gitTools = new McpClient({
    transport: new StdioClientTransport({
      command: "uvx",
      args: ["mcp-server-git", "--repository", "."],
    }),
  });

  // Register tools by name
  const orchestrator = await createOrchestrator({
    directory: "./sops",
    tools: {
      filesystem: fileSystem,
      git: gitTools,
    },
  });

  const result = await orchestrator.invoke(
    "Review the code in src/index.ts and commit any improvements"
  );

  console.log(result);

  // Clean up MCP connections when done
  await fileSystem.disconnect();
  await gitTools.disconnect();
}

main().catch(console.error);

Declaring Tools in SOPs

Agents specify which tools they need in their frontmatter. Only those tools get injected—no context pollution:

---
name: code-reviewer
description: Reviews code and suggests improvements
type: agent
tools:
  - filesystem
  - git
---
# Code Reviewer

## Overview

You review code files and provide actionable feedback.

## Steps

### 1. Read the Code

Use the filesystem tools to read the target file.

**Constraints:**
- You MUST use the `read_file` tool to access file contents
- You MUST NOT guess at file contents because you need accurate information

### 2. Commit Improvements

If changes are needed, use git to commit them.

**Constraints:**
- You MUST write clear commit messages
- You SHOULD make atomic commits for each logical change

The orchestrator can also declare tools in its frontmatter if it needs direct access to MCP capabilities.

Available MCP Servers

| Server | Purpose | Install | |--------|---------|---------| | @modelcontextprotocol/server-filesystem | File system operations | npx -y @modelcontextprotocol/server-filesystem <path> | | mcp-server-git | Git operations | uvx mcp-server-git --repository <path> | | @modelcontextprotocol/server-postgres | PostgreSQL queries | npx -y @modelcontextprotocol/server-postgres <connection-string> | | @modelcontextprotocol/server-sqlite | SQLite database | npx -y @modelcontextprotocol/server-sqlite <db-path> | | awslabs.aws-documentation-mcp-server | AWS documentation | uvx awslabs.aws-documentation-mcp-server@latest |

Browse more at MCP Servers.

HTTP-Based MCP Servers

For remote MCP servers using HTTP transport:

import { McpClient } from "@strands-agents/sdk";
import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/streamableHttp.js";

const remoteTools = new McpClient({
  transport: new StreamableHTTPClientTransport(
    new URL("https://mcp.example.com/sse")
  ),
});

const orchestrator = await createOrchestrator({
  directory: "./sops",
  tools: {
    remote: remoteTools,
  },
});