Open Agent Specification (Agent Spec): A Unified Representation for AI Agents (oracle.github.io)

Open Agent Specification (Agent Spec) is a new portable language for describing agentic systems—from single LLM-powered assistants to structured multi-agent workflows. It defines reusable building blocks (agents and flows), composition rules, and a standardized config format so designers can declare behavior, tools, and inter-agent coordination independently of any execution engine. The spec emphasizes portability, consistency, modularity, and extensibility: agents written once can be run across compatible runtimes, tested side-by-side, and extended with custom components, improving reproducibility and predictable behavior across frameworks. Technically, Agent Spec treats agents as autonomous LLM-based assistants and flows as node-based sequences that implement business logic or multi-step tasks. A Python SDK, PyAgentSpec, generates JSON/YAML configs and integrates with Agent Spec–compatible runtimes such as WayFlow; adapters let other frameworks like AutoGen and LangGraph execute spec-defined agents. By acting as an intermediate representation that complements protocols like MCP and A2A, Agent Spec makes evaluation, comparison, and cross-platform deployment of agentic solutions far simpler—accelerating development, testing, and production portability (pip install "pyagentspec==25.4.1").