OpenAI Agents SDK

pip install "agentguard-oss[openai-agents]"

guarded_tool registers a function with the Guard and returns an Agents SDK FunctionTool.

from agents import Agent, Runner

from agentguard import Guard
from agentguard.adapters.openai_agents import guarded_tool

guard = Guard("policy.yaml", audit="agentguard.jsonl")


def fetch_page(url: str) -> str:
    """Fetch a documentation page."""
    ...


agent = Agent(
    name="Researcher",
    instructions="Answer using the documentation.",
    tools=[guarded_tool(guard, fetch_page, capability="network.request")],
)
result = Runner.run_sync(agent, "What does asyncio.TaskGroup do?")

Options

Argument	Default	Meaning
capability	required	The tool's capability
name, description	function name, docstring	What the model sees
return_denials	True	Return denials as {"error": ..., "reasons": [...]}; False raises, and the SDK reports a tool error
**options		Passed to guard.tool: executor, verifier, url_arg, ...

Already decorated with @guard.tool? Use from_guarded(guarded_function). Async functions are supported.

The Agents SDK also has its own guardrails and needs_approval. They work alongside AgentGuard; AgentGuard's policy, risk checks and audit apply to every call regardless.