Agent governance

Agents

Make Codex, ChatGPT, MCP, A2A, and service agents useful without letting them bypass permissions, approvals, or audit trails.

1 agent principals Open agent console

Controlled interfaces

Agent access model

3 items

Agents should use documented APIs and tools, never private internals or raw provider credentials.

OpenAPI Normal service integration and structured auth/RBAC checks happen through REST/OpenAPI.

MCP Direct agent tools, prompts, and resources stay allowlisted and permission checked.

A2A and CloudEvents Agent delegation and async events must preserve actor, scope, approval, and audit context.

Permissions

3 items

Agentic work is helpful only when it is scoped and reviewable.

Tool permissions Tool calls should map to RBAC permissions and risk tiers.

Approval gates Agents may prepare high-risk changes but must not self-approve them.

Audit evidence Record what the agent requested, which actor authorized it, and what changed.

existing tools