Mandate

Mandate is an authorization and execution control plane for autonomous financial agents.

AI agents can already discover markets, choose trades, and sign actions. The missing layer is proving that a specific action is actually within the authority granted by the human or institution behind that agent. A valid agent signature proves identity; it does not prove permission.

Mandate lets a principal define bounded economic authority once—capital limits, allowed agents, assets, representations, venues, recipients, expiry, and portfolio-wide constraints. Agents can then reason and propose actions independently, while Mandate deterministically verifies every proposal before execution.

The core model is:

Agents propose. Mandate authorizes. Markets settle.

Mandate enforces both agent-local permissions and principal-global portfolio limits. It uses reservations and replay protection so multiple agents cannot independently consume the same authority, binds approved actions to exact execution parameters, and fails closed when required constraints cannot be represented safely. A compromised or mistaken agent remains a subset of the principal’s authority.

The current build includes multi-agent portfolio mandates, the Mandate Room for resolving competing allocations, reusable bounded V3 delegation, durable execution/reconciliation, a security firewall demonstrating VALID AGENT ≠ VALID ACTION, and a thin @mandate/sdk that lets external agents integrate the same authorization boundary without duplicating protocol logic.

We also demonstrate a confirmed autonomous execution on Robinhood Chain Testnet, with the principal authorizing the mandate once and no additional per-trade wallet approval for the demonstrated V3 execution. The canonical settlement used 64 MDUSD → 6.4 MDEMO through the Mandate execution gate.

These are valueless demo assets. This is not an NVDA trade and not a Robinhood Stock Token. The live transaction proves the authorization, execution, replay, reservation, and reconciliation path—not production stock trading.

Mandate’s broader goal is simple: one authority layer for many agents across many financial markets.

Progress During Buildathon During the buildathon, Mandate evolved from an intent-verification concept into a working end-to-end authorization and execution system for autonomous financial agents. We built the core authority model, portfolio-wide accounting, multi-agent allocation controls, replay-safe reservations, the Mandate Room for resolving competing agent requests, and a V3 delegated execution flow where a principal can authorize bounded behavior once and supported agent actions can execute without a new wallet approval for every trade. We also deployed the V3 execution gate to Robinhood Chain Testnet and completed a confirmed autonomous fixture settlement through the full path: principal authorization → agent proposal → Mandate verification → reservation → execution → reconciliation. The resulting transaction and execution evidence are exposed in the demo, alongside a security firewall showing that the same valid agent is refused when it attempts an unauthorized recipient, venue, representation, or amount. The final build also includes a durable recovery/reconciliation layer, a thin @mandate/sdk for external agent integrations, a complete developer documentation site, adversarial and invariant testing, and a production-style web demo. The buildathon version proves the core thesis: A valid agent is not automatically authorized to perform a valid action. Mandate gives autonomous agents freedom to act, while keeping economic authority bounded by the principal.

Mandate

视频

技术栈

描述

本次黑客松进展