Sladk Agents

Your company's Slack AI agents of the new era of generative AI powered by Google Agent Development Kit (ADK)

---

📑 Table of Contents

• 🎯 The Problem
• 💡 The Solution
• 🏗️ Architecture
• ✨ Key Features Implemented (ADK Concepts)
• 🚀 Setup & Installation
• 🎮 Usage
• 🖥️ Demos
• 🔧 Extending the Agent
• 📊 Impact & Value
• 🛣️ Future Roadmap
• 📚 References
• 📜 License
• 👤 Author
• 🙏 Acknowledgments

---

🎯 The Problem

Enterprise teams rely heavily on Slack for daily collaboration—sharing updates, asking questions, coordinating projects, and making decisions. However, accessing AI-powered assistance typically requires:

• Context switching: Leaving Slack to use external AI tools (Gemini web, chatGPT, etc.)
• Manual copy-paste workflows: Moving conversation context between platforms
• Lost productivity: Teams lose time switching between tools
• Fragmented knowledge: AI insights don't stay where the work happens

While Slack recently introduced AI Agents as a native feature, building production-ready agents still requires significant integration work, especially when leveraging advanced AI frameworks.

💡 The Solution

Sladk Agents bridges the gap between Slack's native AI Agent capabilities and Google's powerful Agent Development Kit (ADK). It provides a ready-to-deploy framework that enables enterprises to:

• Deploy AI agents directly within Slack's native UI (side panel, threads, mentions)
• Leverage Google ADK's multi-agent architecture with sub-agents for specialized tasks
• Maintain conversation context through session management
• Extend functionality with built-in and custom tools
• Stream responses in real-time for a seamless user experience

Key Value Proposition: Teams can now access sophisticated AI assistance exactly where they collaborate—no context switching, no copy-paste, just intelligent help when and where they need it.

🏗️ Architecture

┌─────────────────────────────────────────────────────────────────────────────┐
│                              SLACK WORKSPACE                                │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────┐ │
│  │  Side Panel │  │   Threads   │  │  @mentions  │  │   Direct Messages   │ │
│  │  Assistant  │  │             │  │             │  │                     │ │
│  └──────┬──────┘  └──────┬──────┘  └──────┬──────┘  └──────────┬──────────┘ │
└─────────┼────────────────┼────────────────┼────────────────────┼────────────┘
          │                │                │                    │
          └────────────────┴────────────────┴────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                         SLADK AGENTS (Bolt for Python)                      │
│  ┌────────────────────────────────────────────────────────────────────────┐ │
│  │                          AsyncApp + Socket Mode                        │ │
│  │  • Real-time WebSocket connection to Slack                             │ │
│  │  • Async handlers for non-blocking operations                          │ │
│  │  • Native streaming support (chat_stream)                              │ │
│  └────────────────────────────────────────────────────────────────────────┘ │
│                                    │                                        │
│                                    ▼                                        │
│  ┌────────────────────────────────────────────────────────────────────────┐ │
│  │                           LISTENERS                                    │ │
│  │  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐                  │ │
│  │  │  Assistant   │  │    Events    │  │   Actions    │                  │ │
│  │  │  • started   │  │  • mentions  │  │  • feedback  │                  │ │
│  │  │  • message   │  │              │  │              │                  │ │
│  │  └──────────────┘  └──────────────┘  └──────────────┘                  │ │
│  └────────────────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                         GOOGLE ADK (Agent Layer)                            │
│  ┌────────────────────────────────────────────────────────────────────────┐ │
│  │                         LlmAgent (Root Agent)                          │ │
│  │  • Model: Gemini 2.5 Flash (configurable)                              │ │
│  │  • Custom system instructions                                          │ │
│  │  • Safety settings & temperature control                               │ │
│  └────────────────────────────────────────────────────────────────────────┘ │
│                    │                    │                    │              │
│                    ▼                    ▼                    ▼              │
│  ┌──────────────────────┐ ┌──────────────────────┐ ┌──────────────────────┐ │
│  │    SearchAgent       │ │     CodeAgent        │ │   Custom Tools       │ │
│  │  (Sub-Agent)         │ │   (Sub-Agent)        │ │                      │ │
│  │  ┌────────────────┐  │ │  ┌────────────────┐  │ │  ┌────────────────┐  │ │
│  │  │ google_search  │  │ │  │ Code Executor  │  │ │  │  get_weather   │  │ │
│  │  └────────────────┘  │ │  └────────────────┘  │ │  │  (wttr.in)     │  │ │
│  └──────────────────────┘ └──────────────────────┘ │  └────────────────┘  │ │
│                                                    │  ┌────────────────┐  │ │
│                                                    │  │  + extensible  │  │ │
│                                                    │  └────────────────┘  │ │
│                                                    └──────────────────────┘ │
│  ┌────────────────────────────────────────────────────────────────────────┐ │
│  │                        SESSION MANAGEMENT                              │ │
│  │  • InMemorySessionService: Maintains conversation state per thread     │ │
│  │  • Context Compaction: Efficient token usage (interval=5, overlap=1)   │ │
│  │  • User/Session ID mapping: Slack thread_ts → ADK session              │ │
│  └────────────────────────────────────────────────────────────────────────┘ │
│  ┌────────────────────────────────────────────────────────────────────────┐ │
│  │                              RUNNER                                    │ │
│  │  • Async execution with run_async()                                    │ │
│  │  • Streaming response chunks to Slack                                  │ │
│  └────────────────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                            GEMINI API                                       │
│                     (Generative AI Backend)                                 │
└─────────────────────────────────────────────────────────────────────────────┘

✨ Key Features Implemented (ADK Concepts)

This project demonstrates 5 key concepts from the Google ADK course:

1. Multi-Agent System

• Root Agent (LlmAgent): Orchestrates user requests and delegates to specialized sub-agents
• SearchAgent: Handles web search queries using Google Search tool
• CodeAgent: Executes Python code for calculations and data processing

2. Tools Integration

• Built-in Tools: google_search for real-time web information
• Code Execution: BuiltInCodeExecutor for running Python code
• Custom Tools: get_weather() function demonstrating extensibility

3. Sessions & State Management

• InMemorySessionService: Maintains conversation context across messages
• Session persistence: Each Slack thread maps to a unique ADK session
• User isolation: Sessions are scoped per user and thread

4. Context Engineering

• EventsCompactionConfig: Optimizes token usage with:
  • compaction_interval=5: Compact every 5 events
  • overlap_size=1: Maintain context continuity

5. Effective Use of Gemini

• Model: Gemini 2.5 Flash (configurable via environment)
• Safety Settings: Configurable content filtering
• Streaming: Real-time response streaming to Slack

🚀 Setup & Installation

Prerequisites

• Python 3.10+
• A Slack workspace with admin permissions
• Google API key with Gemini access

Tip

You can obtain a Google API with Gemini in the Google AI Studio.

1. Clone the Repository

git clone https://github.com/jonigl/sladk-agents.git
cd sladk-agents

2. Create Slack App

Tip

You can also review the original Slack Bolt template instructions here.

1. Go to api.slack.com/apps/new → "From an app manifest"
2. Select your workspace
3. Paste contents of manifest.json → Click Next → Create
4. Click Install to Workspace → Allow

3. Configure Environment

cp .env.sample .env

Edit .env with your credentials:

# Slack Configuration
SLACK_APP_TOKEN=xapp-...        # Basic Info → App-Level Tokens → connections:write
SLACK_BOT_TOKEN=xoxb-...        # OAuth & Permissions → Bot User OAuth Token

# Google ADK Configuration
GOOGLE_API_KEY=your_google_api_key
AGENT_MODEL=gemini-2.5-flash    # or gemini-2.0-flash, gemini-1.5-pro
AGENT_APP_NAME=Sladk_App
AGENT_NAME=Sladk_AI_Agent

# Optional: Custom system instruction
DEFAULT_SYSTEM_INSTRUCTION="You're a helpful assistant in a Slack workspace..."

4. Install Dependencies

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

5. Run the Agent

python3 app.py

Or using Slack CLI:

slack run

---

🎮 Usage

Configure your Slack Preferences

1. Open Slack → Click your profile picture → Preferences
2. Go to Navigation → Go to the App agents & assistants
3. Click the checkbox for Show app agents

sladk-agents-configure-slack-workspace.mp4

Interact with the Agent

Once running, interact with your agent in three ways:

Method	How
Side Panel	Click the agent icon in Slack's top right corner to open the AI assistant
@Mention	Type @Sladk AI Agent in any channel
Direct Message	Send a DM to the bot

Example Interactions

• "Search for the latest news about AI agents" → Uses SearchAgent
• "Calculate the compound interest on $1000 at 5% for 10 years" → Uses CodeAgent
• "What's the weather in San Francisco?" → Uses custom get_weather tool

🖥️ Demos

Demo 1

• Using google search built-in with ToolAgent / subagent
• Using a custom tool to get the weather
• Interacting with slack thread messages and mentions

sladk-agents-demo-thread-custom-tool-mention-google-search.mp4

Demo 2

• Python code execution using dedicated subagent as a tool

sladk-agents-demo-code-execution.mp4

🔧 Extending the Agent

Adding Custom Tools

Create new tools in ai/tools/custom_tools.py:

def my_custom_tool(param: str) -> str:
    """
    Description of what this tool does.
    Args:
        param: Description of parameter
    Returns:
        str: Description of return value
    """
    # Your implementation
    return result

Then add to the agent in ai/llm_caller.py:

from ai.tools.custom_tools import get_weather, my_custom_tool

agent = LlmAgent(
    # ...
    tools=[get_weather, my_custom_tool, AgentTool(agent=search_agent)],
)

📊 Impact & Value

Metric	Improvement
Context Switching	Eliminated - AI lives in Slack
Response Time	Real-time streaming responses
Tool Access	Web search, code execution, custom tools in one place
Session Memory	Maintains context across conversation
Deployment	Single Python app, no infrastructure complexity

🛣️ Future Roadmap

• Memory Bank Integration: Long-term memory across sessions
• MCP Tools: Model Context Protocol for external integrations
• Agent Engine Deployment: Cloud Run / Agent Engine deployment
• Observability: Tracing and metrics with OpenTelemetry
• A2A Protocol: Agent-to-agent communication for complex workflows

📚 References

• Google ADK Documentation
• Slack AI Apps Documentation
• Bolt for Python
• Gemini API

📜 License

This project uses dual licensing:

Component	License	File
Base code (Slack Bolt template)	MIT License	LICENSE
Sladk Agents additions (agent logic, architecture, ADK integration)	CC-BY-SA 4.0	LICENSE_CAPSTONE

The base Slack Bolt template code is covered by the MIT License (original work by Slack Technologies, LLC). All new agent logic, architecture, and Google ADK integration created for this Capstone project is licensed under CC-BY-SA 4.0 as required by the Kaggle competition rules.

👤 Author

Jonathan Gastón Löwenstern - GitHub: @jonigl

Built with ❤️ for the Kaggle & Google: Agents Intensive - Capstone Project - November 2025

🙏 Acknowledgments

• Slack for the bolt-python-assistant-template foundation
• Google for the Agent Development Kit (ADK)
• The 5-Day AI Agents Intensive Course with Google community for inspiration and guidance