Vertex AI Agent Builder
Google Cloud's managed agent platform using the Agent Development Kit (ADK). Supports unlimited hierarchy depth via sub-agent trees — the concept fit is good, though provisioning is code-first (Python) rather than REST.
Overview
Vertex AI Agent Builder uses Google's Agent Development Kit — a Python-first framework for defining multi-agent systems. Agents are defined as Python classes that extend ADK base types, with sub-agents registered as tools. The hierarchy can be arbitrarily deep via nested sub-agent trees.
Because ADK is code-first, the Envelope integration will use a serialiser that generates ADK Python code from the Envelope team definition, then deploys it via the Vertex AI deployment API. Similar in approach to the LangGraph integration.
Concept mapping
| Envelope concept | Vertex AI equivalent |
|---|---|
| Team | Agent engine / multi-agent system |
| Agent | ADK Agent |
| Run | Agent invocation |
| Install | Agent deployment via Python ADK |
| reportsToKey | Sub-agent tree in ADK |
| requiredSecrets | Google Secret Manager / env vars |
Provisioning approach
Why code-first provisioning — Vertex AI ADK does not expose a REST API for agent provisioning — agents are defined and deployed via Python code. The Envelope integration must generate valid ADK Python from the team definition and submit it for deployment.
This is the same challenge as the LangGraph integration and requires similar depth of work in the serialiser layer. The upside is that ADK's sub-agent tree supports arbitrary hierarchy depth — matching the Envelope schema without approximation.
Integration status
- [x] Research complete
- [x] Concept mapping confirmed
- [ ] ADK serialiser implementation
- [ ] Conformance testing
- [ ] General availability