adk/

Eino ADK Examples

This directory provides examples for Eino ADK:

• Agent
  • helloworld: simple hello-world chat agent.
  • agent/ralph-loop: autonomous iteration loop with filesystem-backed work state and verification gates.
  • intro
    • chatmodel: example about using ChatModelAgent with interrupt.
    • custom: shows how to implement an agent which meets the definition of ADK.
    • workflow: examples about using Loop / Parallel / Sequential agent.
    • session: shows how to pass data and state across agents by using session.
    • transfer: shows transfer ability by using ChatModelAgent.
    • agent_with_summarization: shows how to add conversation summarization to long-running agents.
    • http-sse-service: exposes ADK Runner as an HTTP SSE service.
  • human-in-the-loop
    • 1_approval: sensitive-operation approval.
    • 2_review-and-edit: review and edit tool-call arguments.
    • 3_feedback-loop: iterative writer/reviewer feedback loop.
    • 4_follow-up: ask follow-up questions when required information is missing.
    • 5_supervisor: supervisor pattern with approval.
    • 6_plan-execute-replan: plan-execute-replan with review/edit.
    • 7_deep-agents: Deep Agents with follow-up.
    • 8_supervisor-plan-execute: nested supervisor and plan-execute agents with interrupt.
  • multiagent
    • plan-execute-replan: basic example of plan-execute-replan agent.
    • supervisor: basic example of supervisor agent.
    • layered-supervisor: another example of supervisor agent, which set a supervisor agent as sub-agent of another supervisor agent.
    • integration-project-manager: another example of using supervisor agent.
    • deep: Deep Agents example.
    • integration-excel-agent: ADK-integrated Excel Agent.
  • agentic
    • research_assistant: example of using AgenticModel with ADK typed agent, server-side search, local tools, native filesystem middleware, streaming events, and AgenticMessage output.
    • retry_max_output_tokens: retries when model output is truncated by max_output_tokens.
  • cancel
    • graceful-exit: cancel and resume nested agents safely from terminal signals.
  • middlewares
    • skill: loads Agent skills from filesystem.
    • dynamictool/toolsearch: dynamically retrieves and injects relevant tools from a large tool set.
  • common
    • tool/graphtool: wraps Graph/Chain/Workflow as Agent tools.
    • model, prints, store, trace: shared helpers used by examples.

Additionally, you can enable coze-loop trace for examples, see .example.env for keys.