AWS is open-sourcing a new markdown format called Agent SOPs, designed to make it easier to build AI agents and learn from the shortcomings of its earlier model-driven agent development approach.

Hyperscalers and other vendors, including AWS, have been promoting LLM-driven agent development as a faster way for enterprises to scale agents in production workloads, as it uses an LLM’s reasoning to generate a workflow for agents to follow, in contrast to a developer having to write hundreds of lines of custom code to define a workflow.

Earlier this year, AWS open-sourced an SDK called Strands Agents that it used internally to build agents using LLMs.

However, AWS claims that during its internal use of Strands Agents, its developers encountered issues while deploying agents built with the SDK.

The SDK’s reliance on model-driven reasoning, according to AWS, often produced unpredictable outcomes once agents hit production workloads, leading to inconsistent results, misinterpreted instructions, and high-maintenance prompt engineering — all of which were impediments to adoption at scale.

To bypass these challenges and yet avoid writing lines of custom code, AWS came out with Agent SOPs or Standard Operating Procedures, which are standardized natural language instructions combined with RFC 2119 keywords, such as “MUST”, “SHOULD”, and “MAY”, that offer a way for developers to guide agents to generate a desired workflow.

Essentially, the instructions, parameters, and keywords in SOPs create a structure that acts as a scaffold around which the agent thinks, ensuring that it generates the desired workflow.

During internal use, AWS stated that its teams successfully utilized the SOPs to perform tasks ranging from code reviews and documentation generation to incident response and system monitoring, without needing to write complex custom code to generate workflows.

Building on that success, the hyperscaler has released the code and repositories for Agent SOPs on GitHub, allowing other developers to adopt the same patterns for their own use cases.

Adoption of SOPs, according to AWS, becomes easier as the markdown format works across LLMs, vibe coding platforms, and other agentic frameworks.

“Agent frameworks like Strands can embed SOPs as system prompts, development tools like Kiro and Cursor can use them for structured workflows, and AI models like Claude and GPT-4 can execute them directly,” AWS executives wrote in a blog post.

Further, the hyperscaler said that SOPs can be chained together to execute complex, multi-phase workflows.