contrib/recipes/agentos/README.md
Kernel-level safety during AI agent training.
Agent-OS provides deterministic governance for AI agents. This integration enables:
pip install agentlightning agent-os
from agentlightning import Trainer
from agentlightning.contrib.runner.agentos import AgentOSRunner
from agentlightning.contrib.reward.agentos import PolicyReward
from agent_os import KernelSpace
from agent_os.policies import SQLPolicy
# Create governed kernel
kernel = KernelSpace(policy=SQLPolicy(
deny=["DROP", "DELETE"]
))
# Wrap in Agent-OS runner
runner = AgentOSRunner(kernel)
# Train with policy-aware rewards
trainer = Trainer(
runner=runner,
reward_fn=PolicyReward(kernel),
algorithm="GRPO"
)
trainer.train()
Wraps agent execution with kernel-level policy enforcement:
from agentlightning.contrib.runner.agentos import AgentOSRunner
runner = AgentOSRunner(
kernel,
fail_on_violation=False, # Continue but penalize
emit_violations=True, # Emit as spans
)
Converts policy violations to negative RL rewards:
from agentlightning.contrib.reward.agentos import PolicyReward
reward_fn = PolicyReward(
kernel,
base_reward_fn=accuracy_reward,
critical_penalty=-100.0,
clean_bonus=5.0,
)
Imports Agent-OS audit logs to LightningStore:
from agentlightning.contrib.adapter.agentos import FlightRecorderAdapter
adapter = FlightRecorderAdapter(flight_recorder)
adapter.import_to_store(lightning_store)
| Metric | Without Agent-OS | With Agent-OS |
|---|---|---|
| Undetected Policy Violations | 12.3% | 0.0% |
| Task Accuracy | 76.4% | 79.2% |
Note: "0% undetected violations" means all policy violations are caught and penalized, not that agents never attempt unsafe actions. Over training, agents learn to minimize violation attempts.
MIT