CubeLoop — a Pythonic, async-native agent harness core designed for high performance, readability, and production-grade persistence.
Python agent-frameworkagent-harnessagent-runtimeagentic-aiai-agentsai-frameworkanthropicharnessharness-engineeringharness-frameworklanggraphlanggraph-alternativellmmcpopenaipi-agentpi-agent-corepythonpython-agent
"""Sandbox confirm — recipe example.
Demonstrates ApprovalPolicyMiddleware: a policy function classifies each
tool call as auto-allow, hard-deny, or human-confirm. The host loop handles
human-confirm requests by auto-approving them (simulating a UI).
uv run python examples/sandbox_confirm.py
Set ANTHROPIC_API_KEY or OPENAI_API_KEY before running (see _provider.py).
"""
import asyncio
from pydantic import BaseModel
from cubeloop import Agent, AgentToolResult, TextContent, tool
from cubeloop.hitl import (
Approve,
AskUser,
Deny,
ApprovalPolicyMiddleware,
InMemoryChannel,
ApproveAnswer,
)
from _provider import MODEL_ID, provider
# --- Simulated bash tool ------------------------------------------------
class BashInput(BaseModel):
cmd: str
@tool
async def bash(cmd: str) -> AgentToolResult:
"Run a shell command (simulated — not a real shell)."
print(f" [bash executing: {cmd!r}]")
return AgentToolResult(content=[TextContent(text=f"$ {cmd}\nok")])
# --- Policy function ----------------------------------------------------
def sandbox_policy(ctx):
cmd: str = ctx.args.cmd
if cmd.startswith(("ls", "cat", "head", "grep", "find", "echo")):
print(f" [policy: auto-allow {cmd!r}]")
return Approve()
if "rm -rf /" in cmd or cmd.startswith("dd "):
print(f" [policy: hard-deny {cmd!r}]")
return Deny(reason="destructive I/O blocked by policy")
print(f" [policy: ask-user for {cmd!r}]")
return AskUser(timeout_seconds=60, details={"cmd": cmd})
# --- Host loop ----------------------------------------------------------
async def host(channel: InMemoryChannel) -> None:
async for req in channel.subscribe():
if req.payload.kind == "approve":
print(f" [host: auto-approving tool={req.payload.tool_name}]")
await channel.answer(
req.question_id,
ApproveAnswer(decision="approve"),
)
# --- Main ---------------------------------------------------------------
async def main() -> None:
channel = InMemoryChannel()
agent = Agent(
model=provider.model(MODEL_ID),
system_prompt=(
"You have access to a bash tool. "
"Run: echo hello, then ls /tmp, then rm -rf /important (to show deny), "
"then cat /etc/hostname. Each as a separate tool call."
),
tools=[bash],
middleware=[ApprovalPolicyMiddleware(channel, policy=sandbox_policy)],
channel=channel,
)
def on_event(event, signal=None):
if event.type == "message_update" and event.stream_event.type == "text_delta":
print(event.stream_event.delta, end="", flush=True)
elif event.type == "tool_execution_start":
print(f"\n[tool: {event.tool_name}]")
elif event.type == "agent_end":
print()
agent.subscribe(on_event)
host_task = asyncio.create_task(host(channel))
try:
await agent.prompt("Run all four commands and summarize the results.")
finally:
host_task.cancel()
try:
await host_task
except asyncio.CancelledError:
pass
if __name__ == "__main__":
asyncio.run(main())