Research note · Research
Scoped Authority for Financial Agents
An agent should act inside limits you set, not with your keys. This note maps the design space for delegated authority in financial agents.
Agent-native finance implies agents that trade, pay and execute strategies. The naive implementation hands an agent a key with full balance access — autonomy in name, a single point of failure in practice. Scoped authority is the counter-proposal: delegated permission defined by action allowlists, per-action and aggregate limits, asset and contract scopes, expiry windows and instant revocation, with human approval thresholds above configured amounts.
The interesting questions are not cosmetic. Where does the policy live — on-chain for auditability, or local for privacy? What proves an agent stayed inside its mandate after the fact? What is the revocation latency a user should be able to rely on? Each answer trades auditability against the very context privacy this project exists to research.
The smallest honest artifact for this note is a local simulator: configure an agent, run scripted actions, watch allows, denials and approval gates, then revoke. It proves the interaction design, not security.
What this note covers
- Session keys, per-action limits and expiry windows
- Allowlists for assets, contracts and actions
- Human approval thresholds and immediate revocation
- Proof that the agent stayed inside its mandate