Vibe Coding with Confidence — Itamar Friedman, Qodo

Key Concepts

• Gen 3.0 AI for Developers: Moving beyond autocomplete and chat-based AI to command-line interface (CLI)-driven agentic workflows.
• Vibe Coding vs. Coding with Confidence: Shifting from rapid, potentially low-quality code generation to reliable, high-quality software development with AI assistance.
• AI Across the SDLC: Integrating AI throughout the software development lifecycle, not just within the IDE.
• Agentic Workflows: Utilizing AI agents that can perform end-to-end tasks and communicate with each other.
• MCP (Multi-Agent Collaboration Protocol) & A2A (Agent-to-Agent Communication): Enabling agents to work together, either through pipelines or parallel communication.
• Holistic AI Solutions: Addressing all aspects of software development with integrated AI tools, similar to cloud security solutions.
• CLI as the New Interface Leader: Leveraging the command-line interface for its flexibility, background execution capabilities, and workflow integration.

AI for Developers: From Autocomplete to Agentic Workflows

The speaker, Itar Friedman, CEO and co-founder of Codto, discusses the evolution of AI in software development, differentiating between "noobs" and "enterprise" developers.

• Gen 1.0 (Autocomplete): Initial AI tools focused on code completion within IDEs. While helpful, they weren't game-changers.
• Gen 2.0 (Chat-Based AI): Tools like ChatGPT allowed even junior developers to generate more code, but raised concerns about code quality and review burden for senior developers.
• Gen 3.0 (CLI-Driven Agentic Workflows): The future involves using a command-line interface (CLI) to give commands to AI agents that act as team members, performing end-to-end flows across the SDLC.

Vibe Coding vs. Coding with Confidence

The speaker contrasts "vibe coding" with "coding with confidence."

• Vibe Coding: Rapid code generation, often prioritizing speed over quality and maintainability.
• Coding with Confidence: Achieving 10x-100x development speed while maintaining high quality, reliability, and trust through AI-powered workflows.

The speaker references Andrej Karpathy's initial enthusiasm for vibe coding and his subsequent rethinking, emphasizing the need for context, quality, and maintainability. Karpathy's suggested workflow involves manual context integration, which the speaker argues isn't a true game-changer.

AI Across the SDLC and Agentic Workflows

The speaker emphasizes the importance of integrating AI across the entire software development lifecycle (SDLC), not just within the IDE.

• End-to-End Flows: For "noobs," this might mean generating a fully managed software from a prompt. For enterprise, it involves reliable and versatile agentic workflows.
• Trust and Quality: Agentic workflows must include high-quality reviewing and testing capabilities to ensure trust and reliability.

The speaker breaks down developer tasks into planning, writing code, testing, and reviewing, highlighting that "vibe coding" primarily focuses on initial planning and code writing. He argues that a true game-changer involves AI assisting with all these tasks, especially testing and reviewing, to prevent breaking things.

Shifting Left and Holistic Solutions

The speaker discusses "shifting left" review and testing, meaning performing these tasks earlier in the development process.

• Squeezing the V: The goal is to compress the traditional "V" shape of software development (planning -> coding -> testing -> review) by integrating testing and review earlier.
• MCP & A2A: This involves using Multi-Agent Collaboration Protocols (MCP) and Agent-to-Agent (A2A) communication to enable different agents to work together, either in pipelines or in parallel.

The speaker draws an analogy to cloud security, where holistic solutions like Wiz replaced individual security applications for each part of the cloud. He argues that a similar holistic approach is needed for AI in software development.

CLI as the New Interface Leader

The speaker argues that the command-line interface (CLI) is becoming the new interface leader for AI-powered development.

• Flexibility and Workflow Integration: CLIs can be run in the background, integrated into different workflows, and piped together.
• Simon's Example: The speaker references Simon Willison's talk, noting that he used a CLI for demonstrations, not an IDE.

The speaker introduces Codto's new CLI tool, which allows developers to:

• Chat with AI agents.
• Call specific agents (e.g., "explain agent").
• Create new agents with specific goals and instructions.
• Utilize agents with built-in tools (e.g., Codto Cover agent for automatic coverage generation).

The speaker emphasizes the ability to pipe agents together, creating complex workflows. He also mentions the potential for A2A communication, where agents run in parallel and communicate with each other.

Codto's Multi-Agent Architecture

The speaker briefly describes Codto's multi-agent architecture, which includes:

• Deep Research Agent: For in-depth code analysis.
• Codto Merge: A code review tool that collects best practices over time.
• Codto Aware: An IDE tool that shifts left by bringing context and best practices to the code writing process.
• CLI Tool: For workflow automation and agent management.

Demo and Future Vision

The speaker provides a brief demo of the Codto CLI tool, showing how to create and call agents. He emphasizes the ability to integrate best practices and success/failure criteria into agent workflows.

The speaker concludes by envisioning a future with a "swarm of agents," each specializing in different tasks and having different credentials and best practices. He predicts that this future is coming quickly, around 2025-2026.

Conclusion

The key takeaway is that AI in software development is evolving beyond simple code completion and chat-based assistance. The future lies in CLI-driven agentic workflows that integrate AI across the entire SDLC, enabling developers to achieve significant productivity gains while maintaining high quality and reliability. Codto's CLI tool and multi-agent architecture are designed to facilitate this transition.