Agent SOP defines standard operating procedures (SOPs) for AI agents as markdown files. Each SOP states an objective, takes parameters, and lists steps whose constraints use the RFC 2119 keywords MUST, SHOULD, and MAY, so an agent follows a repeatable process while still reasoning about each step.
The project ships a library of built-in SOPs and a package that distributes them in several forms: as Python modules usable as a Strands Agents system prompt, as prompts served by an MCP server, as Cursor commands, and as Agent Skills in the SKILL.md format. You can add your own SOPs alongside the built-in ones, and an external SOP with the same name overrides the built-in one.
| SOP | Purpose |
|---|---|
| codebase-summary | Analyzes a codebase and generates documentation |
| pdd | Prompt-driven development for problem solving and system design |
| code-task-generator | Breaks requirements into planned code tasks |
| code-assist | Implements code with a test-driven Explore, Plan, Code, Commit workflow |
| eval | Plans and runs agent evaluations with the Strands Evals SDK |
The prompt-driven development SOPs write their output to a shared folder structure organized by what is worth committing: summaries, planning documents, code tasks, and transient scratch notes. SOPs can also instruct an agent to record its progress so a run can resume after a failure.
Features
- RFC 2119 constraints: each step carries MUST, SHOULD, and MAY constraints that control agent behavior without scripting it
- Parameterized inputs: required and optional parameters turn a single-use prompt into a reusable template
- Built-in SOP library: codebase-summary, pdd, code-task-generator, code-assist, and eval workflows
- MCP server: exposes SOPs as prompts that MCP clients discover and invoke on demand
- Agent Skills output: generates SKILL.md skills for Claude and other skill-compatible agents, also published as a Claude Code plugin marketplace
- Cursor commands: converts SOPs into Cursor slash commands that prompt for required parameters
- Strands Agents integration: SOPs import as Python modules and serve directly as an agent's system prompt
- Custom SOPs: load your own SOP files alongside the built-in set, with external SOPs overriding built-ins of the same name
- SOP authoring: an agent-sop-author skill and a format rule let a coding agent write and validate new SOPs from a description
- Progress tracking: SOPs can tell an agent to document progress so an interrupted run can resume
