Ra-AID SUPER AGENTIC Coder : This 3-STEP Opensource Agentic CODER Beats CLINE & CURSOR?

Raid AI Coder Summary

Key Concepts: Autonomous AI coder, Lang graph, agent-based task execution, three-stage architecture (research, planning, implementation), token consumption, Dart AI, API key configuration, human-in-the-loop mode, chat mode, web research, expert tool, model configuration, AER integration.

1. Introduction to Raid AI Coder

Raid is a new open-source autonomous AI coder built on Lang graph's agent-based task execution framework. It can optionally integrate with AER. It aims to autonomously develop software. Its key differentiator is its three-stage architecture:

• Research Stage: Analyzes code bases and gathers context.
• Planning Stage: Breaks down tasks into specific, actionable steps.
• Implementation Stage: Executes each plan step sequentially.

The speaker expresses concern about the potential for high token consumption due to this multi-stage workflow.

2. Dart AI: A Project Management Tool

The speaker promotes Dart AI, an AI-powered project management tool, highlighting its features:

• Generating entire project plans from simple prompts.
• Automatically detecting duplicate tasks.
• Creating subtasks.
• Assigning tasks to the AI for completion (e.g., generating blog posts, researching topics, creating thumbnails).

Dart AI offers integrations with GitHub (linking issues and PRs), Slack, Discord, and an API for custom workflows. It's free for teams up to four people, with paid plans offering more features.

3. Raid Installation and Setup

Installation is done via pip install raid-ai. API keys are required for the chosen provider (default is Anthropic Claude). API keys must be exported as environment variables. Examples:

• ANTHROPIC_API_KEY for Anthropic Claude models.
• OPENAI_API_KEY for OpenAI models.

Options exist for Open Router, Deep Seek, and Gemini providers. Different API keys can be set for the "expert tool" used for complex reasoning.

4. Raid Usage Patterns and Features

Raid has various usage patterns, controlled by command-line flags:

• Basic usage: raid -m "task description"
• Research-only mode: raid --research_only -m "task description"
• Human-in-the-loop mode: raid --hil -m "task description" (allows the agent to ask questions).
• Chat mode: raid --chat -m "task description" (interactive chat interface).

Web Research: Raid can autonomously search the web for information using the Tav API. This is enabled by setting the TAVILY_API_KEY. Tavily has a free tier.

Expert Tool Configuration: The --expert_provider flag allows specifying different providers for the expert tool (e.g., OpenAI 3.5 Turbo, Deep Seek R1).

Model Configuration: Different models can be configured for different stages of the workflow (e.g., a more powerful model for planning, a faster model for implementation).

AER Integration: The --use_aer flag integrates Raid with AER, leveraging AER's agentic capabilities for code generation. AER will automatically select its model based on available API keys (Claude if Anthropic is set, otherwise OpenAI). AER's behavior can be configured using the AER_FLAGS environment variable.

5. Testing and Evaluation

The speaker tests Raid on an image Cropper tool application, tasking it with adding an option to overlay another image.

Results:

• The task consumed approximately 600,000 tokens, deemed excessive.
• The implementation was functional but lacked the ability to move the overlay image.
• The speaker found the implementation unsatisfactory given the high token consumption.

6. Speaker's Perspective and Recommendations

The speaker doesn't recommend using Raid with already capable models like Claude. He suggests it might be more suitable for less capable models like Gemini (potentially using the free Gemini API) or local models, where the recursive, agentic approach could compensate for the model's limitations.

7. Conclusion

Raid AI coder holds potential, particularly for use with less capable models. However, its high token consumption and complex usage patterns are drawbacks. A more simplified and user-friendly design would be beneficial.