Workshop 5: DeepAgents & Automation Duration: 60 minutes | Level: Advanced | Prerequisites: Workshops 1-4
What You’ll Master DeepAgents transforms your expert from a chatbot into an autonomous agent that can complete complex, multi-step tasks.
Understanding DeepAgents
Learn how autonomous agents differ from standard chatbots
Workspace Management
Create and manage persistent file systems
Task Automation
Build and track task lists for complex workflows
Code Execution
Enable your expert to write and run code
HITL Control
Configure human-in-the-loop for safety
What are DeepAgents? DeepAgents are AI agents that can:
┌─────────────────────────────────────────────────────────────────┐ │ DEEPAGENT CAPABILITIES │ ├─────────────────────────────────────────────────────────────────┤ │ │ │ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ │ │ Tasks │ │ Files │ │ Code │ │ │ │ Management │ │ System │ │ Execution │ │ │ └─────────────┘ └─────────────┘ └─────────────┘ │ │ │ │ │ │ │ ▼ ▼ ▼ │ │ Create todos Read/write Run Python, │ │ Track progress files JavaScript, │ │ Multi-step Persistent Shell commands │ │ workflows storage │ │ │ │ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ │ │ Memory │ │ Search │ │ Browser │ │ │ │ System │ │ Web │ │ Automation │ │ │ └─────────────┘ └─────────────┘ └─────────────┘ │ │ │ │ │ │ │ ▼ ▼ ▼ │ │ Remember context Research topics Navigate websites │ │ across sessions Find information Extract data │ │ │ └─────────────────────────────────────────────────────────────────┘
Standard Chat vs DeepAgents Feature Standard Chat DeepAgents Conversations ✅ ✅ File Access ❌ ✅ Task Management ❌ ✅ Code Execution ❌ ✅ Multi-step Workflows Limited ✅ Persistent State Limited ✅ Autonomous Actions ❌ ✅
Enabling DeepAgents Mode Activate in Chat
Open Settings
Click the settings icon in the chat top bar
Enable Deep Mode
Toggle “Deep Mode” or “DeepAgents Mode”
View Workspace
The workspace drawer appears on the right side
The Workspace Interface
Tasks Panel View and manage todos:
See all created tasks
Track completion status
Add manual tasks
Delete or modify
Files Panel Access the file system:
Browse created files
Edit content
Download files
Upload new files
Task Management Task Statuses Status Icon Meaning Pending ⚪ Not started yet In Progress 🔵 Currently working Completed 🟢 Finished successfully Failed 🔴 Encountered errors Cancelled 🟠 Manually cancelled
Creating Tasks You can create tasks by:
Asking the Agent : “Create a task to analyze the sales data”Natural Language : “I need you to do X, Y, and Z”Manual Addition : Click “Add Task” in the workspace
Task Hierarchy DeepAgents supports nested tasks:
📋 Main Project ├── 📋 Research Phase │ ├── ✅ Gather requirements │ ├── 🔵 Analyze competitors │ └── ⚪ Summarize findings ├── 📋 Development Phase │ ├── ⚪ Design architecture │ └── ⚪ Implement features └── 📋 Testing Phase └── ⚪ Write tests
🎯 Exercise: Task Workflow Try this conversation with your DeepAgent:
You: I need to create a marketing report. Can you help me plan this out? Agent: I'll create a structured plan for your marketing report. Let me set up the tasks: 📋 Tasks Created: 1. ⚪ Define report objectives and target audience 2. ⚪ Research market trends and competitors 3. ⚪ Gather internal performance data 4. ⚪ Create outline and structure 5. ⚪ Write executive summary 6. ⚪ Design visualizations 7. ⚪ Review and finalize Should I start with the first task?
File System Operations Available File Operations
Read Files Access and analyze existing files "Read the contents of report.md"
Write Files Create new files or update existing "Create a file called notes.txt with..."
Search Files Find content across files "Search for 'revenue' in all documents"
Edit Files Modify specific parts of files "Update the introduction section of report.md"
File Types Supported Type Extensions Use Cases Text .txt, .md, .json Notes, configs, data Code .py, .js, .ts, .html, .css Scripts, websites Data .csv, .json Datasets, exports Docs .md, .html Reports, documentation
🎯 Exercise: File Operations
Create a File
Ask: “Create a markdown file called project-plan.md with a basic outline for a mobile app project”
Read and Modify
Ask: “Read project-plan.md and add a timeline section”
Create Related File
Ask: “Create a tasks.json file listing the tasks from the project plan”
Code Execution Enabling Code Execution DeepAgents can execute code in isolated sandboxes:
Code execution is powerful but requires caution. Always review code before approving execution in HITL mode!
Supported Languages
Python Full Python 3.x support with common libraries
JavaScript Node.js runtime for JS/TypeScript
Shell Bash commands for system operations
Code Execution Examples Data Analysis
File Processing
API Integration
You: Analyze this CSV data and create a summary Agent: I'll write Python code to analyze the data: [Creates and runs Python script] Results: - Total records: 1,245 - Average value: $847.32 - Top category: Electronics (34%)
You: Convert all the markdown files to HTML Agent: I'll create a script to batch convert: [Creates and runs conversion script] ✅ Converted 5 files: - intro.md → intro.html - chapter1.md → chapter1.html ...
You: Fetch the latest weather for New York Agent: I'll query the weather API: [Executes API call] Current weather in New York: 🌡️ Temperature: 72°F ☁️ Conditions: Partly cloudy 💨 Wind: 8 mph
E2B Workspace Templates DeepAgents uses E2B (Execution to Browser) for code execution:
Template Use Case Features Python Data Science Analysis, ML pandas, numpy, matplotlib Node.js Web dev, APIs npm packages, TypeScript Next.js + Tailwind Full-stack web React, Tailwind CSS Vite + React Frontend apps Fast builds, HMR
Human-in-the-Loop (HITL) What is HITL? HITL mode requires human approval before the agent takes actions:
┌─────────────────────────────────────────────────────────────────┐ │ HITL FLOW │ ├─────────────────────────────────────────────────────────────────┤ │ │ │ Agent Plans Action │ │ │ │ │ ▼ │ │ ┌─────────────────────────────────────────┐ │ │ │ "I want to create 3 files and │ │ │ │ execute this Python script. │ │ │ │ Should I proceed?" │ │ │ │ │ │ │ │ [Approve] [Reject] [Modify] │ │ │ └─────────────────────────────────────────┘ │ │ │ │ │ ▼ │ │ Human Reviews and Decides │ │ │ │ │ ▼ │ │ Agent Executes (if approved) │ │ │ └─────────────────────────────────────────────────────────────────┘
HITL vs Auto Mode Mode Behavior Best For Auto Agent acts immediately Trusted tasks, speed HITL Asks for approval Sensitive operations, learning
Configuring HITL
Open Settings
Click settings in the chat interface
Toggle HITL
Enable “Human in the Loop” mode
Set Thresholds (Optional)
Configure which actions need approval
🎯 Exercise: HITL Workflow
Enable HITL mode
Ask the agent to create a complex project structure
Review each proposed action
Approve, reject, or modify as needed
Observe how the agent adapts to your feedback
Complex Workflow Example Building a Web App with DeepAgents Let’s walk through a complete workflow:
Define the Project
You: Create a simple todo app using React and Tailwind CSS
Agent creates task list:
⚪ Set up project structure
⚪ Create React components
⚪ Style with Tailwind
⚪ Add functionality
⚪ Test and refine
Execute Tasks
The agent works through each task, creating files: 📁 Created files: ├── index.html ├── App.jsx ├── components/ │ ├── TodoList.jsx │ ├── TodoItem.jsx │ └── AddTodo.jsx └── styles/ └── app.css
Test the Result
Agent runs the preview server: 🚀 Preview available at: https://sandbox-xxx.e2b.dev
Iterate
You: Add dark mode support
Agent updates the code and shows results Best Practices
Start with HITL Enable HITL when learning or working with new workflows
Break Down Tasks Complex projects work better with clear subtasks
Review Generated Code Always check code before running in production
Use Templates E2B templates provide stable, tested environments
Advanced Features Memory Integration DeepAgents remember context across sessions:
Project state is preserved
Preferences are learned
Previous decisions inform future actions
Multi-Agent Collaboration Combine DeepAgents with team members:
TechSupport AI (DeepAgent) ├── ResearchBot (gathers information) ├── CodeHelper (writes and reviews code) └── DocWriter (creates documentation)
Scheduled Automation Set up recurring DeepAgent tasks:
Daily reports
Weekly summaries
Automated data processing
🎯 Challenge Exercise Complete this comprehensive project:
Create a Data Dashboard
Ask DeepAgent to create a web dashboard showing sales data
Include Charts
Add visualizations using a charting library
Make it Interactive
Add filters and date range selection
Document
Have the agent create README documentation
Key Takeaways
Autonomous Power DeepAgents can complete complex multi-step tasks independently
Persistent Workspace Files and state persist across conversations
Code Execution Write, run, and iterate on code in real-time
Safe Control HITL mode provides oversight for critical operations
Next Steps 
