Skip to content

manifesto-compliance.md

Latest commit

 

History

History
49 lines (35 loc) · 2.42 KB

manifesto-compliance.md

File metadata and controls

49 lines (35 loc) · 2.42 KB
Raw

AI Agent Manifesto Compliance

The Cabin Crew Protocol is the Reference Implementation for the AI Agent Manifesto.

This document details how the protocol strictly adheres to the 5 Immutable Laws + the Capability Isolation principle.

1. The Principle of Separation

"Reasoning must be separate from action. Intent must be decoupled from Execution."

Implementation:

  • Flight Plan Mode: Engines generate a plan (PlanToken) in flight-plan mode. No state changes occur.
  • Take-Off Mode: Execution happens only in take-off mode, which mandates a valid PlanToken.
  • Preflight Check: The Orchestrator validates the plan before authorizing execution.

2. The Principle of Provenance

"Every output must carry its creation story... cryptographic link between the agent's identity, the input context, the model used, and the final output."

Implementation:

  • PlanToken: Contains workspace_hash, model (e.g. gpt-4), engine_id, and protocol_version.
  • Binding: This token is cryptographically bound to the artifacts it produced.

3. The Principle of External Sovereignty

"No agent can be its own authority. Safety guardrails must exist outside the agent's cognitive loop."

Implementation:

  • Gateways: LLMGateway and MCPGateway wrap all external calls, enforcing policy (rate limits, DLP) outside the engine logic.
  • Orchestrator Veto: The Orchestrator can deny a PreflightOutput regardless of the Engine's confidence.

4. The Principle of Immutable Evidence

"Logs must be proofs, not just text."

Implementation:

  • Signatures: AuditEvent includes a cryptographic signature field.
  • Hash Chain: AuditEvent supports chain_hash for ledger-style non-repudiation.

5. The Principle of Ephemeral Identity

"Static keys are a failure of architecture."

Implementation:

  • Identity Tokens: The protocol supports identity_token (OIDC/JWT) in EngineInput for short-lived, verifiable workload identity.

6. The Principle of Capability Isolation (Draft)

"Agents must not simultaneously process untrusted input, access sensitive systems, and perform state-changing operations... external supervision is mandatory."

Implementation:

  • Human-in-the-Loop: Explicit support via:
    • PreflightOutput.decision: 'REQUIRE_APPROVAL'
    • WorkflowState: 'WAITING_APPROVAL'
    • ApprovalRequest / ApprovalResponse structures.