adk-product-engineers

I want you to use ADK (https://github.com/google/adk-samples https://google.github.io/adk-docs/ https://docs.cloud.google.com/agent-builder/agent-development-kit/overview) to implement 5 examples of agents that PMs/product engineers may be interested in in 5 different directories within this repo. Each directory should be structured similarly to the official Google ADK samples. We will use Python. For each directory, there should be a README with clear instructions on how to get setup. You should make sure you are not just directly copying 1:1 the ADK samples, but implementing something new that has value. You can take inspiration from https://github.com/Shubhamsaboo/awesome-llm-apps if needed. Here’s a concrete spec for five ADK (Python) agent apps that would be useful to PMs and product engineers, with each app living in its own directory and following the standard ADK “agent project” shape (, , ) so it runs cleanly with and the dev UI (). ([Google GitHub][1]) Proposed repo layout (top-level) I’d mirror the ADK samples convention of a parent “agents” folder that contains multiple standalone agent projects: Each directory should be runnable in isolation via: for CLI testing ([Google GitHub][1]) from the parent folder that contains the agent directories (e.g., ) ([Google GitHub][1]) And each README should include a “Gemini API key” path (simple) plus a “Vertex AI” path (enterprise) since ADK supports Gemini auth via AI Studio or Vertex AI. ([Google GitHub][2]) A shared README pattern for every agent directory To keep these consistent and “sample-like,” each README should have the same sections: 1. What this agent does (one paragraph + 3 example prompts) 2. Architecture overview (agents/tools/memory/artifacts) 3. Setup (venv + + ) 4. Run (CLI + dev UI) 5. Demo data (how to run without any external accounts) 6. Optional integrations (off by default; behind tool confirmations) 7. Evals (how to run regression tests and what they check) This aligns well with ADK’s “agent project” assumptions (directory contains with a , plus for keys). ([Google GitHub][1]) Below are the five app specs. 1. — PRD Studio: from idea → PRD + backlog + risks + success metrics Who it’s for PMs and product engineers who want a structured “requirements → execution-ready” packet that’s consistent across teams. Primary job-to-be-done Turn a fuzzy idea into a reviewable PRD draft, a scoped backlog, and an execution checklist—with explicit assumptions and open questions captured along the way. Core interactions (example prompts) “Here’s a 2-sentence product idea. Turn it into a PRD with goals/non-goals, personas, MVP scope, and success metrics.” “Convert this PRD into user stories with acceptance criteria and edge cases.” “Create a launch readiness checklist and a risk register (privacy, reliability, abuse cases).” “What are the top 10 questions you need answered before engineering starts?” Outputs (PRD draft) (epics/stories/AC) (risk register + mitigations) (north star + guardrails + instrumentation notes) ADK architecture Root: pipeline so the flow is predictable and debuggable (intake → draft → critique → finalize). ([Google GitHub][3]) Sub-agents (LLM agents), each with a tight role: “Intake & Scope Agent” (clarifies, sets non-goals, collects constraints) “PRD Writer Agent” (produces structured PRD) “Backlog Agent” (stories + AC + edge cases) “Risk & Policy Agent” (abuse cases, privacy/security prompts, mitigations) “Review Agent” (finds contradictions, missing metrics, unclear scope) Tools and guardrails Artifact writing tool: saves outputs as ADK artifacts (so users can download versioned PRDs/backlogs). ([Google GitHub][4]) Optional: “Create ticket” tool behind confirmation (disabled by default). Use ADK Tool Confirmation so any write-action pauses for approval before it proceeds. ([Google GitHub][5]) README setup requirements Must document ADK install + venv + with . ([Google GitHub][1]) Must document running with and via . ([Google GitHub][1]) Must include a “demo” path that doesn’t call any external systems (only writes artifacts). Nice “not just a toy” detail Include a built-in PRD template validator: after generating the PRD, the Review Agent checks for missing sections, ambiguous requirements, untestable acceptance criteria, and scope creep, then loops once (max) to patch issues. 2. — Experiment Copilot: design + power + analysis + recommendations Who it’s for PMs, growth engineers, and data-minded product engineers running A/B tests or feature rollouts. Primary job-to-be-done Produce an experiment plan (hypothesis, metrics, segments, power), and then analyze results from a CSV export to recommend “ship / iterate / stop,” plus follow-up experiments. Core interactions (example prompts) “Design an A/B test for improving onboarding completion by 5%. Include power, duration, metrics, and instrumentation events.” “Analyze this results CSV and tell me if we should ship. Explain tradeoffs and guardrails.” “Given a neutral result, propose 3 next experiments and what they would de-risk.” Inputs included in repo Optional: user can point to their own CSV path Outputs (event schema suggestions) (decision + reasoning) ADK architecture (showcases a key ADK constraint + workaround) This sample should explicitly demonstrate ADK’s “one tool per agent” limitation for certain built-in tools, and the recommended approach of wrapping specialized agents using when you need to combine capabilities. ([Google GitHub][6]) Concretely: Root agent: a coordinator that can call two “specialist agents” via : 1. “Stats Code Agent” that uses built-in code execution (only capability) to compute p-values, confidence intervals, uplift, and sanity checks. (It has no other tools.) ([Google GitHub][6]) 2. “Narrative Decision Agent” that converts statistical outputs into a product decision and follow-ups. Tools CSV loader tool (local filesystem) A small set of deterministic helper tools (sample-size calculator, SRM checker) Artifacts for saving the plan + analysis. ([Google GitHub][4]) Evals Include a tiny golden dataset () with 5 scenarios (clear win, clear loss, inconclusive, SRM detected, guardrail regression). Use ADK evaluation criteria for response quality and tool-use trajectory (the agent should actually run the analysis path instead of hand-waving). ([Google GitHub][7]) Add one user-sim scenario like “a PM pushing to ship despite guardrail regression” to test safety/rigor under pressure. ([Google GitHub][8]) README setup requirements Must be runnable with only an API key and the bundled CSV data. Must show how to run the eval suite, with expected pass/fail semantics. 3. — Release Radar: dependency & upstream change impact → upgrade plan + draft PR/issue Who it’s for Product engineers (and PMs who care about release risk) managing upgrades, migrations, and breaking changes across repos. Primary job-to-be-done Given a repo and dependency set, generate a “what changed / does it matter / what’s the migration plan” report. Optionally draft a GitHub issue or PR description—never creating anything without confirmation. Core interactions (example prompts) “Scan my repo’s dependencies (requirements.txt / package.json) and tell me what upgrades are high risk.” “Summarize breaking changes between vX and vY for dependency Z and propose a migration checklist.” “Draft a GitHub issue to track this upgrade, including tasks and owners.” Integrations (high leverage ADK feature) Use to generate GitHub REST API tools from an OpenAPI spec, so you don’t handwrite wrappers for endpoints like releases, compare, issues, and pull requests. ([Google GitHub][9]) Safety and reliability All write operations (create issue / create PR / comment) must be behind Tool Confirmation. ([Google GitHub][5]) Enable the Reflect-and-Retry tool plugin so transient API failures (rate limits, schema mismatches) trigger a structured retry instead of a hard stop. ([Google GitHub][10]) Outputs (ranked upgrades, rationale) (one markdown file per upgrade) (ready to paste, even in “no token” mode) ADK architecture Root LLM agent with tools: OpenAPI-generated GitHub tools (read-only mode if no token) Local parsers for / Artifact writer tool for reports ([Google GitHub][4]) README setup requirements “No GitHub token” mode: only reads public metadata and generates local drafts. “Token mode”: enables issue/PR creation but gated by confirmation. 4. — Voice of Customer Insights: feedback → themes → prioritized roadmap slice Who it’s for PMs who want a weekly/monthly “what users are telling us” report that turns raw feedback into actionable product work, plus engineers who want reproducible triage. Primary job-to-be-done Ingest feedback (CSV/JSON), cluster it into themes, quantify themes (frequency, severity, user segment), and output a prioritized set of recommendations with supporting quotes/examples. Core interactions (example prompts) “Ingest this feedback CSV and give me the top themes, with severity and suggested fixes.” “What changed vs last week? Which themes are trending up/down?” “Turn the top 3 themes into backlog items with acceptance criteria.” Inputs Bundled demo data: containing (timestamp, channel, user_segment, text, optional severity) Outputs (exec summary) (structured themes with labels, examples, counts) (backlog candidates) ADK architecture (makes it feel like software, not chat) Root agent drives a deterministic-ish workflow: 1. “Ingestion step” (tool): load + clean + dedupe 2. “Clustering step” (tool): lightweight theme clustering (e.g., TF-IDF + k-means or embedding cosine groups) 3. “Synthesis step” (LLM): name themes, write insights, propose roadmap slice 4. “Critique step” (LLM): check for overgeneralization and missing negative cases Memory Use ADK to store the previous run’s “themes.json” summary so the agent can answer “what changed since last time?” even across sessions. (Default can be in-memory; optionally document how to use a persistent backend.) ([Google GitHub][11]) Evals Add a small eval set with a fixed demo CSV to ensure: The top theme labels stay stable (avoid random renaming regressions) The agent includes supporting examples and doesn’t fabricate quotes The recommendations match the extracted themes (trajectory consistency) ([Google GitHub][7]) README setup requirements Must run fully offline except for the model call. Must document “memory on/off” behavior, so developers can reproduce results (in-memory resets on restart). ([Google GitHub][11]) 5. — Meeting Ops: agenda + notes + decisions + action tracker + follow-up draft Who it’s for PMs running recurring product reviews, eng leads running execution syncs, and anyone who wants consistent decision logs and action tracking without manual busywork. Primary job-to-be-done Turn a meeting transcript into: a structured summary, a decision log, action items with owners/dates, and a drafted follow-up email + updated tracker. Core interactions (example prompts) “Here’s the transcript. Extract decisions and action items; update the tracker; draft a follow-up email.” “Create an agenda for next week based on last week’s open actions and unresolved topics.” “What decisions did we make about launch scope across the last 3 meetings?” Integrations (ADK Google API toolsets) Use ADK’s Google API tool module toolsets (Calendar/Gmail/Docs/Sheets) so the agent can pull meeting metadata, write notes into a doc, update a sheet tracker, and draft an email. The ADK Python API reference explicitly lists , , , and related Google API tooling. ([Google GitHub][12]) Safety Any action that writes externally (create doc, update sheet, draft/send email) must require Tool Confirmation so the user approves changes before they happen. ([Google GitHub][5]) Artifacts Even with real integrations, also save outputs as artifacts so the sample works in “demo mode” (no OAuth) by writing local versions of: ([Google GitHub][4]) Memory Use to store “team conventions” (how action items are phrased, preferred format, recurring attendees, decision taxonomy) and optionally to answer cross-meeting questions like “what did we decide last month?” ([Google GitHub][11]) Optional: expose as a local service For teams who want to integrate this into another UI, include a short README section showing how to run with ADK’s API server and view the interactive docs at on the running server. ([Google GitHub][13]) README setup requirements “Demo mode” (default): no Google OAuth; uses artifacts/local files. “Google mode” (optional): explains required OAuth setup and scopes at a high level, then shows how to enable the toolsets. One consistency rule across all five samples Every sample should be “useful without external accounts” (just a model key) and then optionally unlock integrations. That keeps these truly runnable as examples, while still demonstrating ADK’s real strengths: multi-agent composition and predictable orchestration via workflow agents, tool integrations (including OpenAPI-generated tools), artifacts/memory, confirmations, and evals. ([Google GitHub][3]) [1]: https://google.github.io/adk-docs/get-started/python/ "Python - Agent Development Kit" [2]: https://google.github.io/adk-docs/agents/models/?utm_source=chatgpt.com "Models & Authentication - Agent Development Kit" [3]: https://google.github.io/adk-docs/agents/workflow-agents/ "Workflow Agents - Agent Development Kit" [4]: https://google.github.io/adk-docs/artifacts/ "Artifacts - Agent Development Kit" [5]: https://google.github.io/adk-docs/tools-custom/confirmation/ "Action confirmations - Agent Development Kit" [6]: https://google.github.io/adk-docs/tools/limitations/ "Tool limitations - Agent Development Kit" [7]: https://google.github.io/adk-docs/evaluate/?utm_source=chatgpt.com "Why Evaluate Agents - Agent Development Kit" [8]: https://google.github.io/adk-docs/evaluate/user-sim/?utm_source=chatgpt.com "User Simulation - Agent Development Kit" [9]: https://google.github.io/adk-docs/tools-custom/openapi-tools/ "OpenAPI tools - Agent Development Kit" [10]: https://google.github.io/adk-docs/plugins/reflect-and-retry/?utm_source=chatgpt.com "Reflect and retry - Agent Development Kit" [11]: https://google.github.io/adk-docs/sessions/memory/ "Memory - Agent Development Kit" [12]: https://google.github.io/adk-docs/api-reference/python/?utm_source=chatgpt.com "Agent Development Kit documentation" [13]: https://google.github.io/adk-docs/api-reference/rest/?utm_source=chatgpt.com "REST API - Agent Development Kit" .... Based on the guide you provided, here is a quick introduction to getting started with Google's Agent Development Kit (ADK) for Python. The ADK is a framework designed to help you build, test, and deploy AI agents that can use tools and interact with users. 1. Prerequisites Python 3.10 or later. pip (package installer). 2. Installation Install the ADK package using pip: 3. Create a Project Generate a new agent project structure automatically: This creates a folder named containing (your main code) and (for configuration). 4. Write Your Agent Code Edit the file to define your agent and its tools. A basic example provided in the docs sets up a simple agent with a custom tool: 5. Set Your API Key You need a Google Gemini API key to run the model. 1. Get a key from [Google AI Studio](https://aistudio.google.com/). 2. Add it to the file inside your project folder: 6. Run the Agent You can run your agent in two ways: Command Line Interface (CLI): Web Interface (for development/testing): (Then visit in your browser)*

Implementation Plan: 5 ADK Agent Projects ✅ COMPLETE [x] Install ADK and verify installation [x] Create directory structure: with 5 subdirectories [x] Implement prd_studio - PRD Studio agent [x] Create agent.py with SequentialAgent pipeline [x] Implement sub-agents (Intake, PRD Writer, Backlog, Risk, Review, Metrics) [x] Add artifact writing tools [x] Create requirements.txt, .env.example, README.md (7.5KB) [x] Fixed and validated successfully [x] Implement experiment_copilot - Experiment Copilot agent [x] Create agent.py with AgentTool pattern [x] Add Stats Analysis Agent and Narrative Decision Agent [x] Include sample CSV data [x] Add eval suite with 5 test cases [x] Create requirements.txt, .env.example, README.md (8.5KB) [x] Fixed and validated successfully [x] Implement release_radar - Release Radar agent [x] Create agent.py with dependency parsing tools [x] Add sample data files (requirements.txt, package.json) [x] Create requirements.txt, .env.example, README.md (8.8KB) [x] Fixed and validated successfully [x] Security fix: Updated Flask from 2.0.1 to 2.3.2 (patched version) [x] Implement voc_insights - Voice of Customer Insights agent [x] Create agent.py with clustering workflow [x] Include sample feedback CSV (20 entries) [x] Add eval suite with 3 test cases [x] Create requirements.txt, .env.example, README.md (9.3KB) [x] Fixed and validated successfully [x] Implement meeting_ops - Meeting Ops agent [x] Create agent.py with transcript parsing [x] Include sample transcript data [x] Create requirements.txt, .env.example, README.md (11.3KB) [x] Fixed and validated successfully [x] Create top-level README with overview and quick start guide [x] Add .gitignore to exclude sensitive files and artifacts [x] Fix all ADK compatibility issues [x] All agents load successfully and are ready to run ✅ [x] Create testing guide (TESTING.md) [x] Create implementation summary (IMPLEMENTATION_SUMMARY.md) [x] Address security vulnerabilities in sample data [x] Enhanced README with detailed venv setup instructions ⭐ Summary Successfully implemented all 5 ADK agent projects with complete documentation: Latest Update Added comprehensive virtual environment setup instructions to main README Includes step-by-step guide for venv creation and activation Platform-specific activation commands (Mac/Linux, Windows CMD, Windows PowerShell) Optional ADK installation verification step Statistics Total Files: 38 files created Total Code: ~3,500+ lines (agents + docs) Documentation: 2,135+ lines across all READMEs Sample Data: 7 files for offline testing Eval Suites: 2 agents with test cases Agents Delivered 1. PRD Studio - Idea → PRD + backlog + risks + metrics (SequentialAgent) 2. Experiment Copilot - A/B test design + analysis (AgentTool pattern) 3. Release Radar - Dependency analysis + migration plans 4. VoC Insights - Feedback clustering + roadmap prioritization 5. Meeting Ops* - Transcript → summary + actions + follow-ups Quality Metrics ✅ All agents load without errors ✅ All agents follow ADK patterns ✅ Complete documentation (45+ KB) ✅ Sample data included ✅ Ready to run with ✅ Compatible with ADK 1.21.0+ ✅ Security vulnerabilities addressed ✅ Comprehensive setup instructions with venv All requirements from the issue have been met! Original prompt This section details on the original issue you should resolve Implement 5 ADK agent projects that can be easily run by newbies I want you to use ADK (https://github.com/google/adk-samples https://google.github.io/adk-docs/ https://docs.cloud.google.com/agent-builder/agent-development-kit/overview) to implement 5 examples of agents that PMs/product engineers may be interested in in 5 different directories within this repo. Each directory should be structured similarly to the official Google ADK samples. We will use Python. For each directory, there should be a README with clear instructions on how to get setup. You should make sure you are not just directly copying 1:1 the ADK samples, but implementing something new that has value. You can take inspiration from https://github.com/Shubhamsaboo/awesome-llm-apps if needed. Here’s a concrete spec for five ADK (Python) agent apps that would be useful to PMs and product engineers, with each app living in its own directory and following the standard ADK “agent project” shape (, , ) so it runs cleanly with and the dev UI (). ([Google GitHub][1]) Proposed repo layout (top-level) I’d mirror the ADK samples convention of a parent “agents” folder that contains multiple standalone agent projects: Each directory should be runnable in isolation via: for CLI testing ([Google GitHub][1]) from the parent folder that contains the agent directories (e.g., ) ([Google GitHub][1]) And each README should include a “Gemini API key” path (simple) plus a “Vertex AI” path (enterprise) since ADK supports Gemini auth via AI Studio or Vertex AI. ([Google GitHub][2]) A shared README pattern for every agent directory To keep these consistent and “sample-like,” each README should have the same sections: 1. What this agent does (one paragraph + 3 example prompts) 2. Architecture overview (agents/tools/memory/artifacts) 3. Setup (venv + + ) 4. Run (CLI + dev UI) 5. Demo data (how to run without any external accounts) 6. Optional integrations (off by default; behind tool confirmations) 7. Evals (how to run regression tests and what they check) This aligns well with ADK’s “agent project” assumptions (directory contains with a , plus for keys). ([Google GitHub][1]) Below are the five app specs. 1. — PRD Studio: from idea → PRD + backlog + risks + success metrics Who it’s for PMs and product engineers who want a structured “requirements → execution-ready” packet that’s consistent across teams. Primary job-to-be-done Turn a fuzzy idea into a reviewable PRD draft, a scoped backlog, and an execution checklist—with explicit assumptions and open questions captured along the way. Core interactions (example prompts) “Here’s a 2-sentence product idea. Turn it into a PRD with goals/non-goals, personas, MVP scope, and success metrics.” “Convert this PRD into user stories with acceptance criteria and edge cases.” “Create a launch readiness checklist and a risk register (privacy, reliability, abuse cases).” “What are the top 10 questions you need answered before engineering starts?” Outputs (PRD draft) (epics/stories/AC) (risk register + mitigations) (north star + guardrails + instrumentation notes) ADK architecture Root: pipeline so the flow is predictable and debuggable (intake → draft → critique → finalize). ([Google GitHub][3]) Sub-agents (LLM agents), each with a tight role: “Intake & Scope Agent” (clarifies, sets non-goals, collects constraints) “PRD Writer Agent” (produces structured PRD) “Backlog Agent” (stories + AC + edge cases) “Risk & Policy Agent” (abuse cases, privacy/security prompts, mitigations) “Review Agent” (finds contradictions, missing metrics, unclear scope) Tools and guardrails Artifact writing tool: saves outputs as ADK artifacts (so users can download versioned PRDs/backlogs). ([Google GitHub][4]) Optional: “Create ticket” tool behind confirmation (disabled by default). Use ADK Tool Confirmat... Fixes addyosmani/adk-product-engineering#1 💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.