Your repo is the agent: OpenGAP's git-native standard

What it does

OpenGAP is a specification and reference CLI that models an AI agent as nothing more than a directory of convention-based files inside a git repository. Drop in agent.yaml for the manifest and SOUL.md for identity, and the repo describes everything from personality and skills to compliance rules and cross-session memory, all without tying you to a specific framework runtime.

The interesting bit

The standard borrows software-engineering rigor that AI tooling often skips. It bakes segregation of duties into DUTIES.md, treats branches as deployment environments, and repurposes git diff and git blame as native audit trails for prompt changes. Agent memory is literally markdown files in a memory/ folder.

Key highlights

• Only two required files (agent.yaml and SOUL.md); everything else is optional.
• First-class compliance patterns including FINRA/SEC-style segregation of duties and role conflict matrices.
• Deterministic SkillsFlow workflows defined in YAML, with explicit step ordering and template data flow—no LLM discretion on execution order.
• Supports monorepo composition where sub-agents in agents/ inherit shared skills, tools, and context from the root.
• Can extract the “identity layer” (prompts, roles, schemas) from existing framework-specific agents like NVIDIA AIQ or CrewAI into this portable format.

Caveats

• The specification is at v0.1.0, and the project has already been renamed twice (gitagent → gapman → opengap), suggesting the interface is still settling.
• While the project claims framework-agnostic export, the README emphasizes porting into OpenGAP from other frameworks more than exporting out; the adapter ecosystem remains unclear.
• Compliance features like SOD validation are exposed through opengap validate --compliance, but the README does not clarify whether enforcement happens at validation time or runtime.

Verdict

Teams building regulated or multi-agent systems who need version-controlled prompts and audit trails should look here. If you just want a quick Python script to call an LLM, this is overkill.