# The "Then & Now" Agent: An Autonomous Content Strategist

*A Capstone Project for the 5-Day AI Agents Intensive Course with Google*

## 1. The Pitch

* **Title:** The "Then & Now" Agent: An Autonomous Content Strategist

* **Subtitle:** A multi-agent system that uses iterative refinement and RAG to transform simple topics into strategic, thought-leadership narratives.

### The Problem

In the fast-evolving tech landscape, content teams struggle to keep up. Manually researching a new concept (like AEO vs. SEO), analyzing its historical context, and drafting a strategic narrative takes days. This bottleneck prevents businesses from establishing thought leadership on emerging trends and answering complex user questions with depth and nuance.

### The Solution

This project is an autonomous agent that acts as a lead Research Analyst and Content Strategist. When given a topic or question via Slack, it uses a multi-agent system to:

1. **Triage & Understand:** Semantically classify user intent to distinguish between casual conversation and a work request.

2. **Research & Read:** Use Google Search and a URL reader tool to gather real-time, external context, including its own long-term memory (RAG).

3. **Synthesize & Refine:** Employ an internal 'Writer-Critic' loop to iteratively draft and refine a strategic 'Then vs. Now' analysis, ensuring high quality before human review.

4. **Collaborate:** Present the refined proposal or direct answer to a human via Slack for a final, conversational approval loop (HITL).

5. **Produce & Learn:** Upon approval, generate the final story, create a Google Doc artifact (via N8N), and ingest the new knowledge into its long-term memory in Firestore.

### The Value

This agent transforms a multi-day research and drafting process into a 5-minute, near-autonomous workflow. It allows a single strategist to produce high-quality, data-grounded thought leadership at scale, enabling businesses to react to market shifts in near real-time. It's not just a writer; it's a strategic partner that can answer questions, draft proposals, and learn over time.

---

## 2. Architecture

This agent is built on a serverless, event-driven microservice architecture to ensure scalability and cost-effectiveness. The core logic resides in Google Cloud, with N8N acting as the "last mile" automation bus for Slack and Google Docs integration.

graph TD

    A\[User in Slack] -->|Message| B(N8N Feedback Workflow);

&nbsp;   B -->|HTTP Request| C\[Dispatcher Cloud Function <br><i>start\_story\_workflow</i>];

&nbsp;   C -->|Triage: SOCIAL| D{Generates Witty Reply};

&nbsp;   D --> E\[N8N Proposal Workflow];

&nbsp;   C -->|Triage: WORK| F\[Cloud Tasks Queue];

&nbsp;   F -->|Async Task| G\[Worker Cloud Function <br><i>process\_story\_logic</i>];

&nbsp;   G --> H{Writer-Critic Loop};

&nbsp;   H --> I\[Gemini API];

&nbsp;   H --> J\[Google Search / Jina Reader];

&nbsp;   H --> K\[Firestore Memory <br><i>agent\_sessions</i>];

&nbsp;   H --> L\[RAG Knowledge Base <br><i>knowledge\_base</i>];

&nbsp;   H --> E;

&nbsp;   E --> M\[Slack];

Loading

* **Slack -> N8N -> Dispatcher:** Handles initial triage (Social vs. Work).

* **Dispatcher -> Cloud Task -> Worker:** The core async pattern to prevent timeouts.

* **Worker:** Orchestrates the research, Writer-Critic loop, and final HITL proposal generation.

* **Feedback Loop:** A separate N8N workflow listens for replies, calling the handle\_feedback\_workflow dispatcher, which in turn uses a worker to process refinements, approvals, or context switches.

---

## 3. Key ADK Features Implemented

This project demonstrates a sophisticated application of **five** core ADK concepts:

✅ **Multi-agent System:**

* **LLM Orchestrator:** The process\_story\_logic and process\_feedback\_logic functions act as orchestrators, making decisions and calling tools.

* **Loop agents:** The "Writer-Critic" while loop is a direct implementation of this pattern for quality refinement.

* **Sequential agents:** The overall workflow (Research -> Loop -> HITL -> Finalize) is a sequential pipeline.

✅ **Tools:**

* **Custom tools:** Built several Python functions for specific tasks (find\_trending\_keywords, create\_euphemistic\_links, critique\_proposal, refine\_proposal).

* **Built-in tools:** Integrated **Google Search**.

* **External Tools:** Integrated Jina Reader, N8N, Slack, and Google Docs, demonstrating complex, real-world interoperability.

✅ **Long-running operations (pause/resume agents):**

* The core HITL workflow is an LRO. The agent **pauses** by saving state to Firestore (logging an adk\_request\_confirmation event) and sending a webhook. It **resumes** when the handle\_feedback\_workflow receives the human's approval, loads the state from memory, and continues.

✅ **Sessions & Memory:**

* **Sessions & state management:** A session\_id tracks every conversation. Agent state (status, topic, event\_log) is explicitly serialized into **Firestore**, which acts as our persistent SessionService.

* **Long term memory (RAG):** The knowledge\_base collection and the ingest-knowledge function are a direct implementation of RAG for the agent's long-term memory.

* **Context engineering:** The event\_log serves as a rich context window for the agent to understand conversational history, enabling advanced features like "Session Graduation" (Social -> Work) and "Context Downgrade" (Proposal -> Question).

✅ **Agent deployment (Bonus):**

* The system is deployed as a multi-function microservice architecture on **Google Cloud Functions (2nd Gen)**, built on Cloud Run. The repository includes a clear, repeatable deployment process using PowerShell scripts.

---

## 4. Setup & Deployment

### Prerequisites

* Google Cloud SDK (gcloud CLI) installed and authenticated.

* Python 3.12+ installed.

* A self-hosted N8N instance.

* A Slack App with Bot Token permissions.

### GCP Setup

1. Create a new GCP Project.

2. Enable the following APIs: Cloud Functions, Cloud Run, Cloud Build, Cloud Tasks, Firestore, Secret Manager, Vertex AI, Custom Search API.

3. Create a Service Account (e.g., aeo-devops-agent-identity) and grant it the following IAM roles:

  * Cloud Tasks Enqueuer

  * Cloud Run Invoker

  * Secret Manager Secret Accessor

  * Vertex AI User

  * Cloud Datastore User

  * Service Account User

4. Create a Firestore database in Native Mode.

5. Create a Cloud Task queue (e.g., story-worker-queue).

6. Create secrets in Secret Manager for google-search-api-key and browserless-api-key.

### Local Setup

1. git clone https://github.com/your-username/google-adk-aeo.git

2. cd google-adk-aeo

3. python -m venv venv

4. .\\venv\\Scripts\\Activate.ps1

5. pip install -r requirements.txt

### Deployment

1. Open PowerShell and run the "Pre-Flight Checklist" to set your environment variables ($project\_id, $search\_id, $N8N\_PROPOSAL\_WEBHOOK\_URL).

2. Execute the gcloud functions deploy commands for the four functions: start-story-workflow, handle-feedback-workflow, process-story-logic, and process-feedback-logic. *Full commands are available in the project documentation or can be provided upon request.*

---

## 5. How to Use

1. Invite the Slack Bot to a channel.

2. Send it a message.

  * For a conversational reply: "Hello!"

  * To get a direct answer: "What is the history of the telephone?"

  * To trigger the "Then vs Now" proposal workflow: "Draft a story about the evolution of video games."

3. Interact with the agent by replying in the thread to refine, approve, or ask follow-up questions.