agentmesh-openai-agents-trust

Trust & governance layer for the OpenAI Agents SDK. Adds policy enforcement, trust-gated handoffs, and tamper-evident audit trails using native SDK guardrails and hooks.

Built by AgentMesh — the open-source trust layer for multi-agent systems. Similar integrations merged into Dify (65K ⭐), LlamaIndex (47K ⭐), and Microsoft Agent-Lightning (15K ⭐).

Install

pip install agentmesh-openai-agents-trust

Quick Start

from agents import Agent, Runner
from openai_agents_trust import (
    trust_input_guardrail,
    policy_input_guardrail,
    GovernanceHooks,
    TrustGuardrailConfig,
    PolicyGuardrailConfig,
    TrustScorer,
    GovernancePolicy,
    AuditLog,
)

# Shared governance state
scorer = TrustScorer()
audit = AuditLog()
policy = GovernancePolicy(
    name="production",
    max_tool_calls=20,
    blocked_patterns=[r"DROP TABLE", r"rm -rf", r"eval\("],
    min_trust_score=0.7,
)

# Create guardrails
trust_config = TrustGuardrailConfig(scorer=scorer, min_score=0.7, audit_log=audit)
policy_config = PolicyGuardrailConfig(policy=policy, audit_log=audit)

# Attach to agents
agent = Agent(
    name="researcher",
    instructions="You are a research assistant.",
    input_guardrails=[
        trust_input_guardrail(trust_config),
        policy_input_guardrail(policy_config),
    ],
)

# Run with governance hooks
result = await Runner.run(
    agent,
    input="Analyze this data",
    run_config=RunConfig(hooks=GovernanceHooks(policy=policy, scorer=scorer, audit_log=audit)),
)

# Verify audit integrity
print(f"Audit entries: {len(audit)}")
print(f"Chain valid: {audit.verify_chain()}")

Trust-Gated Handoffs

from openai_agents_trust import trust_gated_handoff

billing_agent = Agent(name="billing", instructions="Handle billing.")
support_agent = Agent(name="support", instructions="Handle support.")

triage = Agent(
    name="triage",
    handoffs=[
        trust_gated_handoff(billing_agent, scorer=scorer, min_score=0.8),
        trust_gated_handoff(support_agent, scorer=scorer, min_score=0.6),
    ],
)

Features

License

MIT