The Mandrel Specification

The Mandrel Spec is the foundational “Contract” of the Mandrel Project. It is a declarative definition that outlines an agent’s operational boundaries, target MCP tools, security constraints, and lifecycle metadata.

It is designed as a direct extension of the Agent Card standard defined by the Linux Foundation’s Agent2Agent (A2A) project (originally initiated by Google), providing the hardened governance required for enterprise-grade autonomous systems.

The Two States of a Spec

To maintain high-integrity governance, the spec exists in two distinct states:

1. The Source Spec: Authored by the domain developer and stored in version control. This contains the intent of the agent.
2. The Realized Spec: Enriched by the CI/CD pipeline with cryptographic attestations and binary authorization. This is the single source of truth for the Gantry (Operator) and the Metrology Lab (Registry).

Key Concepts

• Core Schema: Detailed breakdown of the schema fields (Governance, Identity, Network, Runtime).
• Identity & Security: Compound identities, SPIFFE, and Nested JWTs.
• Orchestration & Delegation: How agents interact and the Max-Hop protocol.
• Extensions Reference: Detailed field-level reference for Mandrel-specific controls.
• Metrology Contract: The normative contract defining promotion gate rules, JCS canonicalization, and signature verification.

Rationalization with LF A2A

By maintaining strict compatibility with the base A2A Agent Card schema, Mandrel-governed agents remain interoperable with the broader AI ecosystem while exposing high-performance governance extensions enforced by The Collet (Sidecar).

Source

The Mandrel Spec schema and reference YAML are hosted on GitLab: gitlab.com/the-mandrel-project/mandrel-spec