Claude Fable 5 and Agentic Workflows in VS Code

The Evolution of Agentic Workflows

The panel highlights the transition from simple code completion to agentic workflows where models handle multi-step tasks. A key development is the introduction of 'Autopilot' in GitHub Copilot, which is now the default for complex tasks. Unlike previous iterations that relied on simple prompting, Autopilot utilizes a secondary 'utility model' to evaluate task completion, preventing infinite loops by capping execution at three attempts. The panelists note a shift in developer trust, where they now delegate entire tasks to agents and expect them to be completed without constant oversight.

Multi-Model Chaining for Production Quality

To achieve production-grade code, the panelists advocate for a 'battle' approach between models. By chaining models like Claude Opus and GPT-5, developers can force models to review each other's work for accuracy, cleanliness, and edge cases. This iterative loop continues until the models reach a consensus that further improvements offer diminishing returns. While acknowledging this is a token-intensive, brute-force method, the panelists argue it is the most effective way to overcome the 'pathological lying' inherent in current LLMs.

AI-Assisted Utility Development

The discussion showcases how AI lowers the barrier to building custom developer tools. The panelists demonstrate building small, single-purpose utilities (like window resizers and screen capture tools) using Go and web technologies (Wails). They emphasize that the future of development involves using AI to generate these tools on the fly, abstracting away compilation, dependencies, and environment setup. The panel suggests that tools like PowerToys should eventually integrate these AI-generated utilities directly into the OS.

Strategic Model Selection

For efficient workflows, the team recommends a tiered model strategy: using high-performance models (like Claude Fable 5 or Opus) for execution and planning, while relying on smaller, low-latency models for background tasks like generating commit messages or chat titles. This approach optimizes for both cost and speed without sacrificing the quality of the final output.