If you’re someone whose work involves AI automation or a developer, you’ll know that Modern AI agents are powerful but notoriously difficult to guide with consistency. You put the same thing in twice, you get two drastically different results.

There are two ways of driving Modern AI agents:

1. Code defined workflows
2. Fully model-driven agents

Code-defined workflows offer precision but require heavy reengineering to update. Fully model-driven agents offer flexibility but can behave unpredictably.

Agent SOPs (Standard Operating Procedures) provide a middle ground: natural language workflows with structure, constraints, and parameters that help you predict how an AI will perform complex tasks without rigid programming. It was originally developed inside Amazon’s internal builder community, quickly spreading across teams because they made agent behavior both more reliable and easier to evolve. Now, they’re available as an open source technology for everyone.

Why Agent SOPs Were Created?

•  A popular problem point for agentic coding assistance, agents make different decisions for the same inputs: choosing different tools, writing different outputs, formats, or misunderstanding task priorities. Reliability suffers without a predictable structure.

• Creating complex workflows requires deep expertise. Instructions just didn’t transfer well across models and use cases. Changing a prompt that has been used repeatedly often requires long manual evaluation cycles.
• Teams inside Amazon needed a format that was structured enough for consistency but flexible enough to let agents think. That insight led to Agent SOPs.

What Are Agent SOPs?

They are standardized markdown workflows that encode multi-step processes in natural language. Without forcing you to write code, they introduce enough structure to make unpredictable agent behavior predictable and repeatable.

Core features include:

• Structured steps with RFC 2119 keywords
  • Steps use MUST, SHOULD, MAY to add precise behavioral constraints without scripting logic.

• Parameterized inputs
  • SOPs accept parameters so the same workflow can be reused across codebases, teams, and projects.
• Progress tracking and resumability
  • Workflows instruct agents to document steps as they proceed, improving transparency and debugging
• Works everywhere
  • SOPs can run in Strands Agents, Kiro, Cursor, Claude Skills, MCP-compatible tools, or directly in LLMs like ChatGPT and Claude.

Get Started with AWS Strands Agent SOPs

You can get started through:

• GitHub repository: https://github.com/strands-agents/agent-sop
• Installing the package:  pip install strands-agents-sops 
• Running the MCP server:  strands-agents-sops 
• Try it:  @codebase summary 

An Example of Running the Codebase Summary in Strands Agent SOPs

Install the package

 pip install strands-agents-sops 

Register the MCP server

 {
    "mcpServers": {
        "strands-agents-sops": { "command": "strands-agents-sops", "args": [] }
    }
 } 

List available SOPs (via /prompts in Kiro)

•  code-assist 
•  code-task-summary 
•  codebase-summary 
•  pdd 

Run the SOP

 @codebase-summary 

Provie parameters in natural language

 codebase_path: ~/github/agent-sop/python 
 consolidate: true 
 consolidate_target: README.md 

Integrating Strands Agent SOPs into AI Platforms

Strands Agents

You can use SOPs directly as agent system prompts:

 from strands import Agent 
 from strands_tools import editor, shell 
 from strands_agents_sops import code_assist 

 agent = Agent(system_prompt=code_assist, tools=[editor, shell]) 
 agent("Start code-assist sop") 

An example automation use case for this: continuously update README documentation whenever new commits/updates land.

Claude Skills

You can generate skills automatically through:

 strands-agents-sops skills 

For each SOP, it creates a  SKILL.md .

Cursor, Claude Code, and Kiro IDE

You can use SOPs as structured prompts, for steering files, or custom commands.

Python Modules

You can import SOPs directly for custom automation systems.

Authoring Your Own SOPs via AWS Strands Agent SOPs

Creating new SOPs is intentionally dead simple.

Load the SOP format rule

 strands-agents-sops rule 

Describe your workflows in natural language

 I want an SOP that extracts action items, decisions, and follow-up tasks from meeting notes. 

The AI produces a fully structured SOP with parameters, steps, constraints, and output formats.

You can even iterate naturally (e.g. “add deadlines”, “store decisions in a project log”, etc.) until the SOP matches your desired workflow.

Ready to use AWS Strands Agent SOPs?

Agent SOPs are a game changer in the current era of AI automation, allowing you to make the behavior of you AI agents repeatable and predictable. More importantly, they allow you to bridge the gap between different AI workflows (e.g. coded workflows and model-driven agents).

If these features are what you’re looking for to perfect your AI automation workflow, then AWS Agent SOPs are for you.

References

• https://aws.amazon.com/blogs/opensource/introducing-strands-agent-sops-natural-language-workflows-for-ai-agents/
• https://aws.amazon.com/blogs/opensource/strands-agents-and-the-model-driven-approach/