"Kubernetes" for AI Agents
Just as a Kubernetes manifest declares what a container needs and the cluster decides how to run it, an .agf.yaml declares what an agent needs and the runtime decides how to execute it.
The same .agf.yaml file runs on any compliant runtime. No code changes, no vendor lock-in.
Approval workflows, budget limits, and governance policies declared in YAML. Auditors read YAML, not code.
Every .agf.yaml is validated against a JSON Schema. Typos and misconfigurations caught at parse time.
Two-layer governance model lets security teams enforce organization-wide policies on top of agent declarations.
Agent Definitions Are Locked In
Today, defining an agent means writing code in a specific framework. Agent Format provides one portable definition that any runtime can execute.
# LangChain
from langchain.agents import create_react_agent
from langchain_google import ChatGoogleGenerativeAI
from langchain.tools import Tool
llm = ChatGoogleGenerativeAI(model="gemini-2.5-pro")
tools = [search_tool, calculator_tool]
agent = create_react_agent(
llm=llm, tools=tools, prompt=prompt_template
)
# Google ADK
from google.adk import Agent
agent = Agent(
name="analyst",
model="gemini-2.5-pro",
instruction="You are a financial analyst.",
tools=[search_tool, calculator_tool],
)
# PydanticAI
from pydantic_ai import Agent
agent = Agent(
"gemini-2.5-pro",
system_prompt="You are a financial analyst.",
tools=[search_tool, calculator_tool],
)
# AutoGen
from autogen import AssistantAgent
agent = AssistantAgent(
"analyst",
llm_config={"model": "gemini-2.5-pro"},
tools=[search_tool, calculator_tool],
)
# 4 frameworks, 4 incompatible APIs
# Switch runtime = rewrite your agent
metadata:
id: financial_analyst
name: Financial Analyst
interface:
input: string
output: string
action_space:
local_tools:
- alias: search
- alias: calculator
execution_policy:
id: agf.react
config:
instructions: You are a financial analyst.
model: gemini-2.5-pro
# 1 definition → runs on ANY compliant runtime
How It Works
Three steps from definition to execution.
Write a .agf.yaml file describing your agent’s identity, tools, constraints, and execution strategy.
Run agf lint to catch schema errors, semantic issues, and best-practice warnings at authoring time.
Load the YAML into any compliant SDK. The runtime resolves tools, credentials, and policies.
Three Complementary Standards
Agent Format defines what an agent is. A2A and MCP handle how agents communicate and use tools. Together, they form a complete stack for production AI systems.
Agent Format
Declares WHAT the agent is: identity, interface, tools, constraints, governance, and execution strategy.
A2A Protocol
Defines HOW agents communicate: discovery, skill invocation, and task delegation between agents.
Model Context Protocol
Defines HOW agents use tools: tool discovery, invocation, and result streaming via MCP servers.
Towards Foundation-Grade Governance
Agent Format is designed for neutral stewardship — modeled after the OpenAPI Initiative and CNCF projects.
Open Governance
Vendor-neutral Standard Committee with open membership, lazy consensus for minor changes, and supermajority for breaking changes.
Conformance Program
Three-tier badge system — Parse, Execute, and Full Governance — so users know exactly what each runtime supports.
Multi-Language SDKs
Official parsers and validators for Go, Python, Java, and TypeScript. Same schema, same validation, any language.