Agent governance extension for A2A, version 1
What an agent card says about each agent that the A2A protocol itself cannot.
URI: https://quilzo.github.io/spec/agent-governance/v1. This is a proposal with an implementation behind it, namespaced to this project rather than to A2A, not a standard.
Fields, one entry per skill
Every field is read from the enforced declaration, so it describes the gate rather than a promise beside it.
| Field | Meaning |
|---|---|
capabilities | The closed list of operations. Anything else is refused, however the request is phrased. |
autonomy | propose, draft or publish: what the agent may do with what it produces. |
writes | Whether any capability changes stored content, derived from the list. |
humanApproval | Whether a person must agree before anything the agent did becomes public. Always on for publish. |
taintsOnRead | The run is marked once it reads stored content, and a marked run cannot publish itself. |
scope | The content it may read: which ref, types and locales. Empty means unrestricted, stated rather than implied. |
budget | What one run may spend, so a caller can price delegated work before spending anything. |
delegates | The agents it may hand work to, by name. |
hosts | The outside systems it may reach. Empty means none. |
memory | What it retains and for how long; absent when it retains nothing. |
See A2A for how Quilzo publishes and uses it.