Cobo launches Agentic Wallet with Pact guardrails

Singapore-based Cobo launched the Agentic Wallet this week. The product combines a human-agent authorization protocol called a Pact with multi-party computation (MPC) to limit AI agents that perform on-chain transactions.

Cobo built a modular skill framework so developers and users can attach specific capabilities to agents and restrict their actions. Each task an agent runs is wrapped in a dynamically generated Pact that sets execution boundaries and termination conditions. Changhao Jiang, co-founder and CTO, wrote that the Pact “defines what an agent can do, where it must stop, and when execution ends,” and that the rules are enforced at the infrastructure level.

Under the MPC design, signing authority is divided among multiple parties. According to Cobo, a compromised agent, a hallucinating large language model, or leaked credentials alone cannot produce a valid signature to move funds.

The wallet integrates with several AI agent frameworks, including LangChain, the OpenAI Agents SDK, Claude MCP, Agno and CrewAI. It supports more than 80 blockchain networks, among them Ethereum, Base, Arbitrum, Optimism, Polygon and Solana.

The launch comes as other crypto firms develop agent-enabled wallets. Hardware wallet maker Ledger has published an AI-focused security agenda. Coinbase has introduced agentic wallets that allow agents to hold funds, execute trades and perform on-chain operations. Trust Wallet released a toolkit to let agents carry out crypto transactions.

Cobo describes the Pact mechanism as task-specific and time-bound and says MPC splits signing authority so no single compromised component can act alone. The company presented the features as an added safety layer for wallets that delegate actions to AI agents.