Vertex AI Agent Engine hosts Harmonia’s managed ADK hierarchy. It provides cognitive execution and session transport; it does not own durable job state.
How Harmonia uses it
The worker reconstructs a typed invocation snapshot from Firestore, creates a namespaced managed session, invokes one requested specialist, streams state deltas, validates the required output, and deletes the ephemeral session after success or failure.
Each invocation is scoped by workspace, brand, user, job, stage, and operation identifiers. A fresh Agent Engine revision is required when agent schemas, hierarchy, prompts, or Nova skills change.
State and authority
Session state is transport context only. It cannot advance a workflow, approve content, authorize cost, or prove an external effect. Only validated records persisted through deterministic routes have durable meaning.
Failure behavior
Missing resources, malformed state deltas, unavailable sessions, and provider errors fail explicitly. Deployment refuses to invent or silently reuse an unknown managed resource.
Evidence status
The managed-runtime adapter and protocol are locally tested. Deployment scripts express the intended topology; only an authenticated invocation trace proves the configured Agent Engine revision ran.
Relevant files
agent/harmonia_agent/team_runtime.pyagent/harmonia_agent/agent_engine_app.pyagent/harmonia_agent/agents.pyagent/harmonia_agent/agent_engine_deploy.pyinfra/deploy.sh
Last modified on August 28, 2026
