Vertex AI Agent Builder
also known as Gemini Enterprise Agent Platform, Google Cloud Agent Builder
Type: full-code · Vendor: Google Cloud · Language: API (multi-language SDKs: Python, Java, Go, TypeScript via Agent Development Kit) · License: proprietary (managed service); Agent Development Kit is open-source (Apache-2.0) · Status: active · Status in practice: mature · First released: 2024-04-09
Google Cloud's end-to-end platform to build, scale, and govern agents: an open-source Agent Development Kit (ADK) for code-first multi-agent design, Vertex AI Agent Engine as the managed runtime with Sessions, Memory Bank, and Code Execution, plus Agent Garden samples and governance / observability hooks.
Description. Vertex AI Agent Builder (now branded Gemini Enterprise Agent Platform) is Google Cloud's open platform for building, scaling, and governing agents. The build pillar is the Agent Development Kit (ADK) — an open-source framework in Python, Java, Go, and TypeScript that supports both code-defined custom agents and pre-built workflow agents (SequentialAgent, ParallelAgent, LoopAgent), with native multi-agent composition (sub-agents, hierarchical delegation), tools (built-in, custom, MCP, OpenAPI), and the A2A protocol for agent-to-agent communication. The scale pillar is Vertex AI Agent Engine: a managed runtime with Sessions (conversation state), Memory Bank (long-term personalised memory), and Code Execution. Agent Garden ships curated samples for one-click deployment. Governance layers add Cloud Trace observability, IAM-based agent identity, and Security Command Center threat detection.
Agent loop shape. Two-tier shape. ADK defines the agent loop: an LlmAgent reasons over instructions and decides which tool or sub-agent to call, while workflow agents (SequentialAgent / ParallelAgent / LoopAgent) compose deterministic pipelines around LLM agents. Multi-agent hierarchies delegate by transferring control to sub-agents. The composed agent is then deployed onto Agent Engine, which provides the managed runtime, Sessions for conversation state, Memory Bank for persistent long-term memory, sandboxed Code Execution, and Cloud Trace observability. Tools span built-in, custom, OpenAPI, MCP, and A2A endpoints.
Primary use cases
- code-first multi-agent systems built with ADK and deployed on Agent Engine
- workflow agents (sequential / parallel / loop) for predictable pipelines
- long-running production agents with managed Sessions and Memory Bank
- RAG agents grounded by Vertex AI Search and Vertex AI RAG Engine
- agent-to-agent ecosystems over the open A2A protocol
Key concepts
- Agent Development Kit (ADK) (docs) — Open-source agent framework in Python, Java, Go, TypeScript with native multi-agent composition.
- Workflow agents → orchestrator-workers (docs) — SequentialAgent, ParallelAgent, LoopAgent compose deterministic pipelines around LLM sub-agents.
- Agent Engine (docs) — Managed runtime that scales agents in production with Sessions, Memory Bank, and Code Execution.
- Sessions → agent-resumption (docs) — Maintain conversation state across turns; managed by Agent Engine.
- Memory Bank → cross-session-memory (docs) — Persistent long-term memory for user preferences and facts across sessions.
- A2A protocol (docs) — Open agent-to-agent protocol; Agent2Agent integration is a first-class ADK concept.
- Agent Garden (docs) — Curated repository of agent samples, solutions, and tools with one-click deployment.
Patterns this full-code implements —
- ★★Agent Resumption
Agent Engine Sessions maintain conversation state across turns; sessions and memory bank are GA on Agent Engine.
- ★★Agentic RAG
Vertex AI RAG Engine is the first-class retrieval primitive: it indexes private data into a corpus, retrieves relevant context per query, and augments the LLM prompt. ADK integrates RAG Engine and Ve…
- ★★Code Execution
Agent Engine provides managed Code Execution so agents can run Python/JavaScript in a secure, isolated sandbox with state preserved across calls; supports file I/O and TTL up to 14 days.
- ★★Cross-Session Memory
Memory Bank stores long-term user preferences and facts across sessions; explicit Memory Bank API.
- ★★Orchestrator-Workers
ADK ships SequentialAgent, ParallelAgent, and LoopAgent for deterministic orchestration of sub-agents.
- ★★Supervisor
ADK organises agents in a tree (root_agent + sub_agents). Parent LlmAgents delegate to sub_agents via LLM-driven auto-flow (selected by their descriptions) or via a transfer_to_agent tool action.
- ★★Tool Use
ADK exposes a rich tool ecosystem: built-in tools, third-party integrations, and custom code; Google Search grounding included.
- ★★Model Context Protocol
ADK has a dedicated MCP section and ships MCP tool support; MCP tools sit alongside Custom Tools and OpenAPI tools as a first-class tool kind.
- ★★Eval Harness
Evaluation is a first-class ADK area: criteria, user simulation, custom metrics; partner evaluation tools and offline / online monitors.
- ★★Input/Output Guardrails
ADK exposes Callbacks as the design-pattern surface for guardrails; Safety and Security and Responsible AI are documented sections.
- ★★Multi-Model Routing
ADK is model-agnostic with adapters for many providers; can target Gemini, Claude, or other models.
- ★★Structured Output
ADK supports structured output on LlmAgent via output_schema (typically a Pydantic BaseModel). When set, the agent's final response must be a JSON string conforming to that schema; mutually exclusive…
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Alternatives & relatives
Listed as alternative by (2)
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified