Skip to main content
An agent’s capability is not a build-time flag it holds — it is a level granted server-side through the tool profile bound to the agent and the scopes minted into its short-lived token. Raising an agent’s level widens the tool surface it can reach; it never widens the authority policy grants. Every level above Read still clears the policy engine and, for signing, MPC threshold signing.

​
The four levels

All four levels are available now. A level is the tool profile you bind to an agent plus the scopes you request at token exchange — nothing to install, nothing to self-operate.

​
R0 · Read

The read-only floor. The read profile exposes balances, holdings, prices, fee estimates, asset and chain resolution, transaction status and history, and wallet inventory. No signing scope is present, so no signing tool is reachable regardless of policy. R0 is the safe default for reporting and monitoring — see the Reporting & Treasury Insights template.

​
R1 · Prepare

Everything in R0 plus construction of unsigned artifacts: unsigned transfers, swap quotes and simulations, and fee, nonce, and slippage preparation. R1 also covers external-signer and approval handoff — the agent produces a payload for a signer it does not control. Still no wallets:sign, so R1 cannot release a share.

​
R2 · Execute

Adds wallets:sign. The agent drives hosted MPC signing composed with the policy engine: each execution binds to an exact, one-use signing intent, and broadcast is idempotent with finality reconciliation.
The wallets:sign scope enables the tool; the policy engine still decides — a denied intent cannot be signed. See Policy Gates → What a Deny Does.

​
GA · Governed Autonomous

The full governed surface for autonomous operation: per-agent policy, atomic cumulative budgets and reservations, destination/asset/chain/contract/method controls, durable approvals, emergency controls, and execution receipts. The agent operates under its mandate and budget — defined in Agents and Authority and set by an admin via Managing Policy Rules — with durable receipts in the audit trail.

​
What sets an agent’s capability

An agent’s capability is the tool profile an admin bound to it — read, prepare, execute, or governed — set on WalletSuite’s side, never by the caller. The short-lived token the agent connects with can never exceed that profile: a capability the profile does not grant cannot be reclaimed by the model, the prompt, or a downstream call. Chains follow the same rule — an agent can only act on the chains it’s entitled to. See Policy Gates.

Agents & Authority

Identity, wallet grant, mandate, budget, and expiry — the five bound things.

Policy Gates

The deterministic decision point every level above Read passes through.

Tool Reference

The tools each profile exposes.

Token Exchange

How scopes and tool profile reach the agent as a short-lived token.