Architecture Overview

Describes functional boundaries and operational ownership across the agent gateway, control plane, execution plane, and settlement truth layer.

• Identity binds requests to a verified principal or governed runtime.
• Mandates scope what can be proposed and what conditions must be met.
• Policy versions define approval logic and reviewable constraints.
• Independent deterministic risk decides whether execution may proceed.

• Accounts, Pay, and Markets receive deterministic commands only after authorization.
• Agent Gateway acts as the boundary between proposal context and executable financial actions.
• Confidential execution protects sensitive state while preserving common control semantics.
• Private perps extend the execution plane without changing the core trust model.

1. Proposal arrives through the Agent Gateway with idempotency and correlation metadata.
2. Control-plane systems validate identity, mandates, policy, and independent risk.
3. Execution-plane services issue deterministic account, payment, market, or treasury actions.
4. Ledger and settlement publish terminal state, balances, and provenance records.

• Authorization failure: mandate, policy, or risk blocks execution.
• Boundary mismatch: proposal context attempts to bypass governed control surfaces.
• State divergence: client projection differs from canonical ledger or settlement state.

Retriable operations must preserve operation identity, respect replay and rate controls, and never escalate agent authority during recovery.

• Per-boundary dashboards for gateway intake, authorization latency, execution latency, and settlement finality.
• Cross-system trace views keyed by proposal, policy, request, event, and settlement identifiers.
• Ownership-aware incident templates for suspension, replay review, and emergency kill-switch activation.

• RFC Index
• RFC / Technical Specs