Introducing Devid, AI Software Engineer You Can Actually Use

---

---

AI Summary

• Introduction of open source AI software engineer named David
  • Shoutout to Slavic roots
  • ”Actually usable” defined by:
    1. Full access to source code for customization
    2. Trained on real-world coding challenges
    3. Agentic system, not just an assistant
• Agency consists of three agents:
  1. Planner agent
  2. David (developer agent)
  3. Browsing agent
• David’s Workflow:
  1. Creates a plan
  2. Executes tasks such as checking directories and creating files (index.html, style.css, game.js)
  3. FileWriter tool generates files based on requirements
• Demonstration:
  1. Game of Life created and tested successfully
  2. David modifies the game to full screen and removes buttons upon request
• Critique of Devin (another AI software engineer):
  1. Lack of transparency in the chat interface
  2. Prompting required during task execution
  3. Long task completion time (2 hours for a simple task)
• Setting up David:
  1. Clone agency swarm repo
  2. Install Docker for isolation
  3. Build Docker container (Python or Selenium based)
  4. Run container and start coding
  5. Import David using Agency Swarm import command
  6. Full access to source code for customization
• Agency Development:
  1. Planner agent breaks down tasks and directs them
  2. Browsing agent finds and analyzes documentation, exports as files
  3. David reads documentation and executes tasks
  4. Response validators ensure correct agent behavior
  5. Agents communicate as if prompted by user for reliability
• Fine-tuning the Agency:
  1. Update agency manifesto for context and process
  2. Adjust individual agents’ instructions and tools
  3. Use shared state for function execution order
• Conclusion:
  1. Agency can be made autonomous for specific tasks
  2. Encouragement to subscribe to the channel