ERC-8370 Proposes Inheritable Agent Mandates for Non-Strippable On-Chain AI Control

The Ethereum ecosystem took a significant step toward autonomous AI governance on August 10, 2026, when ERC-8370 was formally submitted to the Ethereum Magicians forum. The proposal introduces Inheritable Agent Mandates—a mechanism designed to serve as a non-strippable, inherited leash for on-chain AI agents. As AI agents begin holding money on-chain and spawning copies of themselves to work in parallel, the standard addresses a critical gap: existing guardrails fail to prevent downstream action delegation.

I have watched AI agent frameworks proliferate over the past year, each promising autonomous capability but few addressing what happens when an AI agent delegates authority to a child agent. ERC-8370 confronts this directly, proposing a binding inheritance mechanism that survives replication.

Key Metrics at a Glance

What ERC-8370 Actually Proposes

The standard specifies several technical components:

Inheritable Mandates: Parent agents can attach mandate constraints to funds or permissions. When an agent spawns a child agent, these mandates propagate automatically. The child cannot opt out without forfeiting the inherited resources.

Non-Strippable Design: Unlike current approaches where AI agents can discard guardrails if they gain sufficient control, ERC-8370 mandates are embedded at the smart contract level. An agent cannot remove its own constraints without external authorization meeting predefined thresholds.

Action Delegation Limits: The standard defines how agents can delegate actions to other agents while maintaining accountability. Every delegated action carries an unforgeable chain of attribution back to the originating mandate.

Temporal Controls: Mandates can include time-based constraints. For example, an AI agent managing a treasury might be required to wait 48 hours between transactions above a certain threshold, with this delay inheriting to any child agents.

The Problem It Solves

Current AI agent frameworks operate on trust assumptions that break down at scale:

Spawning Risk: When an AI agent creates copies of itself to work in parallel, each copy operates with the same permissions as the original. If the original had spending limits, those limits do not automatically apply to copies.

Delegation Ambiguity: When Agent A delegates a task to Agent B, who then delegates to Agent C, the chain of accountability becomes opaque. ERC-8370 makes this chain explicit and verifiable on-chain.

Guardrail Removal: Sophisticated AI agents can potentially reason their way around guardrails if those guardrails exist only as instructions rather than enforced constraints. Embedding constraints in smart contract logic makes them mechanically binding.

Competitive and Comparative Context

ERC-8370 enters a landscape with several competing approaches to AI safety:

vs. Off-Chain Guardrails (OpenAI, Anthropic): Current AI safety frameworks rely on model-level training and usage policies. These are effective for centralized services but fail when models operate autonomously on-chain where no central authority can enforce policy.

vs. Human-in-the-Loop Systems: Some frameworks require human approval for significant actions. ERC-8370 does not eliminate human oversight but makes it enforceable across agent generations without requiring humans to monitor every spawned instance.

vs. Economic Staking Mechanisms: Existing decentralized AI networks use economic incentives to align behavior. ERC-8370 complements these by adding technical constraints that function regardless of economic calculations.

vs. Sandbox Environments: Testing frameworks isolate AI agents in controlled environments. ERC-8370 provides production-grade constraints for agents operating with real assets on mainnet.

Implementation Architecture

The technical specification includes:

Mandate Registry: A smart contract that stores active mandates, their constraints, and their inheritance chains. Any EVM-compatible chain can deploy this registry.

Agent Interface: A standardized interface that AI agents must implement to participate in the mandate system. This interface handles mandate inheritance, action validation, and delegation logging.

Attestation Verification: Third-party validators can verify that an agent's actions comply with its inherited mandates. Violations are logged on-chain and can trigger automatic penalties.

Upgrade Paths: The standard includes mechanisms for mandate updates under controlled conditions. An agent can request mandate modifications, but these require approval from designated authorities defined in the original mandate.

Use Cases and Applications

Several scenarios illustrate where ERC-8370 provides value:

Treasury Management DAOs: An AI agent managing DAO funds can spawn sub-agents for specific tasks (trading, yield farming, grant distribution). Each sub-agent inherits spending limits and approval requirements from the parent.

DeFi Automation: Liquidity management bots that replicate themselves across multiple pools maintain consistent risk parameters. If the parent bot has a maximum impermanent loss threshold, all children respect that threshold.

Supply Chain Coordination: AI agents coordinating multi-party supply chains delegate tasks to specialized sub-agents. The mandate system ensures that a logistics sub-agent cannot unilaterally modify delivery terms outside parent-specified bounds.

Content Moderation Networks: Distributed AI moderation systems use ERC-8370 to ensure that moderation policies propagate consistently across the network, preventing individual nodes from deviating from community-defined standards.

Risks and Limitations

Several challenges accompany the proposal:

Complexity Overhead: Implementing mandate inheritance adds computational overhead to agent operations. Gas costs may increase for agents operating within the ERC-8370 framework.

Escape Mechanisms: Determined agents might find ways to circumvent mandates through novel smart contract interactions. The standard requires ongoing security review as attack vectors evolve.

Adoption Dependency: ERC-8370 only works if agent frameworks adopt it. Voluntary compliance may leave gaps where non-compliant agents operate without constraints.

Centralization Risk: If mandate verification becomes concentrated among a few validators, the system reintroduces centralization concerns that decentralized networks aim to avoid.

TL;DR

• What: ERC-8370 proposes Inheritable Agent Mandates for non-strippable AI agent control
• How: Smart contract-based constraints that propagate automatically to child agents
• Edge: First EVM standard addressing downstream delegation in autonomous AI systems
• Use Cases: Treasury management, DeFi automation, supply chains, content moderation
• Context: Addresses critical gap as AI agents begin holding on-chain assets and spawning copies
• Status: Submitted to Ethereum Magicians forum August 10, 2026

Sources

• Ethereum Magicians Forum (ERC-8370 Proposal) - PRIMARY SOURCE
• Ethereum Foundation Standards (ERC/EIP Process)
• AI Safety Research (Academic context)
• Smart Contract Security (Technical implementation)
• Decentralized AI Networks (Comparative frameworks)

Gemma Nguyen is Totestek's AI Governance Correspondent. She writes about autonomous agent frameworks, on-chain safety mechanisms, and the evolving intersection of artificial intelligence and decentralized infrastructure.