Workflow manuals that tell agents what they MUST do

What it does Agent SOPs is a library of markdown-based standard operating procedures that tell AI agents how to execute complex, multi-step tasks—like analyzing a codebase, planning features, or writing code—without drifting off-script. Each workflow is parameterized and uses RFC 2119 keywords (MUST, SHOULD, MAY) to make instructions unambiguous. The package distributes these SOPs as Python modules, MCP server prompts, or Cursor IDE commands.

The interesting bit The project treats agent prompts like infrastructure: versionable markdown files that separate committed documentation from transient scratchpad output in a structured .agents/ directory. It is a rare admission that an AI’s “thought process” should be organized, git-ignored, or committed just like human-generated artifacts.

Key highlights

• Five built-in workflows cover codebase summarization, prompt-driven development, task generation, code assistance, and evaluation.
• RFC 2119 constraint keywords force explicit, unambiguous instructions instead of hand-wavy suggestions.
• Parameterized inputs let the same SOP adapt to different contexts without rewriting the underlying markdown.
• Triple-mode distribution: consume as a Python SDK, an MCP server prompt, or a Cursor IDE slash command.
• Auto-organizes outputs into .agents/ subdirectories based on whether they deserve to be committed or ignored.

Verdict Grab it if you want to stop rewriting prompts for every project and need agents to follow the same playbook twice in a row. Skip it if you prefer to wing every LLM interaction.