Google ADK & Gemini
Four products. One ecosystem.
Google’s agent stack spans interactive terminal work, async GitHub automation, IDE assistance, and full enterprise multi-agent orchestration. For building virtual company agent teams, the Agent Development Kit (ADK) is the centrepiece — Apache 2.0, multi-language, and backed by 100+ prebuilt enterprise connectors.
Agent Development KitGemini CLIJulesCode AssistFour products
Orchestration SDKInteractive CLIAsync GitHub AgentIDE PluginProduct comparison
| Product | Mode | License | Key differentiator |
|---|---|---|---|
| ADK | SDK / library | Apache 2.0 | Graph-based multi-agent orchestration; 100+ SAP/Salesforce/Workday connectors; Vertex AI deployment; A2A protocol |
| Gemini CLI | Interactive terminal | Apache 2.0 | 1M token context window on Gemini 2.5 Pro; generous free tier (60 req/min); multimodal (images, PDFs, video) |
| Jules | Async GitHub agent | Proprietary | Shows a work plan before executing; GitHub issue and PR-native; limited beta access as of mid-2026 |
| Code Assist | IDE plugin | Proprietary | Google Workspace SSO + Cloud Audit Logs; designed for enterprises already on GCP/Workspace |
ADK deep dive — building agent teams
Install ADK and initialise a project
pip install google-adk# or TypeScriptnpm install @google/adk
# Initialise a new agent projectadk create my-agent-teamcd my-agent-teamDefine specialised agents and wire them into a graph
ADK uses a graph model — each node is an agent with a specific role and toolset. Edges define when control passes from one agent to another.
from google.adk import Agent, AgentGraph, tool
research_agent = Agent( name="researcher", model="gemini-2.5-pro", tools=[web_search, read_document], instruction="Research the topic and return structured findings.",)
writer_agent = Agent( name="writer", model="gemini-2.5-flash", tools=[write_file, read_file], instruction="Draft content based on the researcher's findings.",)
graph = AgentGraph()graph.add_edge(research_agent, writer_agent)Connect enterprise systems via prebuilt connectors
ADK ships 100+ connectors for enterprise platforms — no custom integration code required for common systems.
from google.adk.connectors import SalesforceConnector, SAPConnector
salesforce = SalesforceConnector( instance_url="https://myorg.salesforce.com", credentials={"client_id": "...", "client_secret": "..."},)
# The connector exposes typed tools the agent can callcrm_agent = Agent( name="crm_agent", tools=[salesforce.get_opportunity, salesforce.update_deal_stage],)Deploy to Vertex AI for enterprise SLAs
# Deploy the agent graph to Vertex AI Agent Engineadk deploy --project my-gcp-project --region us-central1
# Run locally for developmentadk run --task "Prepare the Q2 sales pipeline report"Pricing
| Product | Free tier | Paid |
|---|---|---|
| ADK + Vertex AI | ADK framework is free (Apache 2.0); Vertex AI has a free trial credit | Vertex AI token pricing — Gemini 2.5 Pro: $1.25/M input · $10/M output (1M+ context: $2.50/$15) |
| Gemini CLI | 60 req/min on Gemini 2.5 Pro with Google account — no API key needed | Google AI Studio or Vertex AI API key for higher limits |
| Jules | Limited beta — task quota applies | Pricing not yet public as of mid-2026 |
| Code Assist | Free tier for individuals | Google Workspace Enterprise pricing for teams with SSO + audit logs |
When to use Google’s agent stack
Enterprise-scale agent teams — SAP, Salesforce, Workday out of the box
Use ADK. If your agent team needs to touch enterprise systems — pull CRM data from Salesforce, raise a purchase order in SAP, update a Workday workflow — ADK’s 100+ prebuilt connectors eliminate weeks of integration work. These aren’t thin wrappers; they handle auth, pagination, and schema mapping so agents can call typed tools rather than raw REST endpoints.
Complex multi-step branching workflows with specialised agents
Use ADK’s graph orchestration. ADK models agent handoffs as a directed graph — a research node passes findings to a writing node, which routes to a review node based on content type. This is more structured than prompt-based orchestration and makes it easier to reason about which agent owns each step, add new branches, and audit execution paths.
GCP data residency and compliance requirements
Use ADK deployed on Vertex AI. All data stays within your chosen GCP region. Google Cloud Audit Logs captures every agent action for compliance reporting. If your operation already runs on GCP and needs GDPR, SOC 2, or ISO 27001 coverage, Vertex AI provides those certifications end-to-end.
Large codebase exploration with a massive context window
Use Gemini CLI. Its 1M token context window can ingest an entire monorepo in one pass — something neither Claude Code nor Pi can do on a standard context budget. This is the right tool for the first exploration of an unfamiliar large codebase: run gemini in the repo root and ask broad questions before committing to a framework for ongoing work.