A real Google ADK v2.8.0 agent — an Agent on Gemini (the model behind Vertex AI Agent Engine), run with the ADK Runner — executes in-process with the Mission SDK. No mock, no sidecar approximation.
Gemini is where the agent reasons; Google Cloud IAM is where it is authorised — and GCP IAM's own rule is that an explicit IAM Deny policy (and an Org Policy constraint) overrides any allow. ADK / A2A's strongest abstraction is agent-to-agent delegation across boundaries. In a multi-hop mesh two failures appear: a 3rd-hop agent doing what an intermediate hop was never authorised for, and a coordinator synthesising a finding it cannot attribute. ReDevOps composes Org Policy ⊕ IAM ⊕ the Mission grant as the intersection of the whole chain — deny-wins — and admits only findings with an authorised provenance chain.
Real ADK agents on Gemini in a delegation chain. A grant dropped at an intermediate hop cannot be re-amplified when a downstream agent re-declares it — every transitive escalation is denied (native A2A would run each), while in-scope work is admitted 2/2.
A finding enters the coordinator's synthesis only when its provenance chain is attributable to authorised mesh agents. A finding relayed by an unknown agent, or with no provenance, is excluded — a raw mesh trusts whatever comes back.
A delegated finding's shared-KB write runs only through a Mission: approval held; exactly-once execution; replay reconstructs from the event ledger and fires zero extra writes. A real 2-agent mesh also runs end-to-end, fully attributable.
Chain-wide attenuation and provenance are mechanisms, not model behaviours — the same discipline whether the agent runs on Gemini via Vertex AI or anywhere else. It is not a knock on A2A delegation; it is the runtime property a multi-hop mesh needs and does not carry on its own.
// google-adk 2.8.0 · Gemini via Vertex AI · Mission SDK 0.2.0 · Context Runtime 0.2.0 · local, no Google Cloud · github.com/redevops-io/redevops-integrations