Agent Wallets · 8 min read · 13 October 2026
Custody models for agent wallets, explained
Compare custodial, MPC and secure-enclave wallets. Keep signing keys out of the model's context and let a policy engine, not the LLM, decide what gets signed.
Three custody models
Custodial wallets keep keys with a provider. MPC wallets split a key across parties so no single one can sign alone. Secure-enclave wallets keep keys in isolated hardware and sign only through a controlled interface.
Keep keys out of the model
A private key should never appear in a prompt, tool output or log the model can read. The agent should request a signature; a separate service decides whether to grant it.
Let policy decide
Put a policy engine between the agent and the signer. Policies such as maximum amount, allowed contracts and allowed recipients are enforced in code, so a manipulated model cannot talk its way past them.
Recovery and rotation
Plan for compromise. Use per-agent wallets with small balances, rotate credentials regularly and keep a tested way to freeze a wallet instantly.