Part of Agent OS - Kernel-level governance for AI agents
Muted agent execution with graph-based constraints.
Mute Agent separates reasoning from execution. A reasoning component proposes actions, and an execution component carries them out — but only if the action passes validation against a constraint graph. Out-of-scope requests return NULL instead of hallucinated responses.
pip install -e .For development with testing tools:
pip install -e ".[dev]"from mute_agent import ReasoningAgent, ExecutionAgent, HandshakeProtocol
protocol = HandshakeProtocol()
reasoning = ReasoningAgent(knowledge_graph, router, protocol)
execution = ExecutionAgent(protocol)
session = reasoning.propose_action(
action_id="read_file",
parameters={"path": "/data/file.txt"},
context={"user": "admin"},
justification="User requested file read"
)
if session.validation_result.is_valid:
protocol.accept_proposal(session.session_id)
result = execution.execute(session.session_id)- Separation of reasoning and execution concerns
- Graph-based constraint validation
- Deterministic action authorization
- Complete audit trail via session tracking
MIT License