Policy playground

Lyra's safety boundary is deterministic code, not a prompt. Configure the fund-control policy and a proposed action below; the verdict is computed by the same pure function that guards the agent on-chain. The model proposes, this disposes.

Policy (AgentPolicy)

Hard cap (SUI/tx)

Auto ceiling (SUI)

Max slippage (bps)

Autonomy

read-only (block all writes)

Coin allowlist (comma-separated; empty = any)

Recipient allowlist (empty = any)

Proposed action

Kind

Amount (SUI)

Asset

Recipient

NEEDS APPROVALdeterministic verdict

Permitted, but material-risk — the agent pauses and asks for human approval before it broadcasts, even under YOLO. The deterministic floor sits beneath the session mode.