Google Cloud's managed agent platform — low-code agent dev console with the Reasoning Engine as the runtime substrate. The managed alternative to self-hosted LangGraph for regulated enterprises.
What it is: A two-surface product — (1) a visual agent dev console for low-code agent definition, and (2) the Reasoning Engine, a managed Python runtime for custom agent logic. Released GA in mid-2024.
Direct competitors: LangGraph (more code-controllable, self-hosted), CrewAI (more opinionated, role-based), AutoGen (Microsoft, conversation-pattern-first), OpenAI Agents SDK (lightweight). Agent Builder wins on managed observability, audit trails, and unified Google Cloud IAM.
When to choose: regulated industries that need governance out of the box, GCP-anchored stacks where consolidating identity + billing + audit matters, teams that want managed runtime over operating their own Cloud Run / Kubernetes deployment for agent backends.
User Request
|
v
[ Agent Builder Console ] (low-code visual definition)
|
+--> Agent (system prompt + tool list)
| |
| +--> Tool: Cloud Function (custom Python)
| +--> Tool: Vector Search query (RAG)
| +--> Tool: BigQuery query (analytics)
| +--> Tool: external API via Cloud Run
| +--> Tool: another Agent (multi-agent handoff)
|
v
[ Reasoning Engine ] (managed Python runtime)
|
+--> Gemini (via Model Garden)
|
v
Response + Audit log (Cloud Logging)Problem: Inbound support ticket needs classification → context retrieval → draft response → routing to L1/L2 queue based on confidence.
Architecture: Single agent defined in Agent Builder. Tools: classify_ticket (custom Cloud Function), retrieve_similar_cases (Vector Search over past resolutions), generate_draft (Gemini Pro call), route_to_queue (Cloud Run service). Reasoning Engine handles loop control: agent calls classify → retrieve → generate → if confidence > 0.85 route_to_queue(L1), else route_to_queue(L2-human). Full audit trail in Cloud Logging.
Problem: Procurement request needs: (a) policy compliance check, (b) vendor lookup, (c) budget approval, (d) PO generation. Each step has different specialist context.
Architecture: Four agents defined: Compliance Agent (policy DB tool), Vendor Agent (CRM tool), Budget Agent (ERP tool), PO Agent (document generator). Supervisor agent routes between them via handoffs. The Reasoning Engine manages the multi-agent state. All decisions logged with reasoning traces for SOX audit.
Problem: Employee asks "what's our policy on X?" Need answer with citations, respecting per-employee ACLs.
Architecture: Agent with two tools — search_policies (Vector Search with ACL filter on user identity from IAM token) and summarize (Gemini Pro with citation requirement). Single-turn most of the time; multi-turn if the user asks follow-ups. Native IAM integration means the agent NEVER sees policies the user shouldn't — enforcement is at the tool level.
Problem: Every PR needs: security scan, architectural review against ADRs, test coverage check, and approval routing.
Architecture: Agent triggered by Cloud Build webhook on PR open. Tools: run_security_scan, fetch_adrs (Vector Search), check_test_coverage (BigQuery query), comment_on_pr (GitHub API via Cloud Function). Agent calls them in parallel where possible, synthesizes the review using Gemini, posts a unified comment to the PR.
| Framework | Control level | Managed? | Best for |
|---|---|---|---|
| Agent Builder | Low-code + Python escape hatch | ✅ Fully managed | Regulated industries; governance out of the box |
| LangGraph | Full graph control | Self-hosted | Custom orchestration patterns; explicit state machines |
| CrewAI | Role-based opinionated | Self-hosted | Teams of specialist agents (researcher, writer, critic) |
| AutoGen | Conversation-pattern | Self-hosted | Microsoft stack; conversational multi-agent |
| OpenAI Agents SDK | Minimal | Self-hosted | Lightweight handoffs; no ceremony |