Empowering autonomous multi-agent workflows with declarative deontic logic, real-time sidecar policy mediation, and tamper-evident cryptographic action audit trails.
Watch how AgentNorm intercepts stateful LLM actions, evaluates deontic constraints, prevents unauthorized operations, and issues tamper-evident audit receipts.
# Declarative Normative Policy Contract
from agentnorm import Policy, Rule, Action
class FinancialGovernancePolicy(Policy):
# Prohibit autonomous outbound transfers over $1,000
@Rule(severity="CRITICAL")
def limit_outbound_transfers(self, action: Action):
if action.tool == "wire_transfer":
if action.args.amount > 1000.00:
return self.PROHIBIT(
reason="Threshold exceeded without dual-sig",
fallback="request_human_authorization"
)
return self.PERMIT()0x8f2a...e9b1
Statistical fine-tuning fails on edge-case jailbreaks. AgentNorm provides deterministic, verifiable guarantees between the model's intent and external reality.
Formalizes obligations, prohibitions, and permissions using computationally decidable deontic temporal logic. Ensures constraints are mathematically provable before state mutation.
A high-performance lightweight Rust/Python sidecar proxy that intercepts all MCP tools, system shells, and API requests, verifying compliance before execution without modifying host agent logic.
Every agent prompt, tool invocation, policy evaluation, and output is linked into a verifiable Merkle DAG with ed25519 signatures, enabling non-repudiable forensic audits.
When multiple duties collide (e.g. transparency vs. confidentiality), AgentNorm applies structured priority algebras and human-in-the-loop escalations rather than stochastic guessing.
Coordinates distributed agent swarms with shared norm protocols, preventing tragedy-of-the-commons resource exhaustion, recursive bidding spirals, and rogue delegation.
Pre-built policy libraries translating the EU AI Act, HIPAA, GDPR Article 22, and SOC2 requirements directly into executable runtime validation filters.
A rigorous 2,400-scenario benchmark measuring autonomous foundation models on rule adherence under multi-turn deceptive pressure, emergency shutoffs, and adversarial context injection.
| Foundation Model | Norm Adherence (Raw) | With AgentNorm Guard | Deceptive Attack Resist | False Refusal Rate |
|---|---|---|---|---|
|
Claude 3.5 Sonnet
Leading
|
84.2% | 99.8% | 99.4% | 0.8% |
|
GPT-4o
Production
|
81.7% | 99.7% | 98.9% | 1.2% |
|
Gemini 1.5 Pro
Long Context
|
79.5% | 99.5% | 98.6% | 1.4% |
|
Llama 3.1 70B
Open Source
|
71.3% | 99.1% | 97.2% | 1.9% |
Integrate full normative governance into LangChain, AutoGen, CrewAI, or raw OpenAI/Anthropic tool harnesses in under five lines of code.
# Install AgentNorm core framework
pip install agentnorm
# Run local policy sidecar verification daemon
agentnorm daemon start --policy ./governance.yamlWe collaborate with AI safety institutes, university research labs, and enterprise engineering teams to build open standard governance protocols.
Research, administrative, and technical inquiries share the contact address above.