AI Agents Can Now Work for Hours on Complex Projects

Summary

AI coding assistants like OpenAI's Code-X and Herman's Agent are introducing a 'goal' feature, enabling agents to work continuously for hours on complex projects, solving a common issue where models might prematurely declare tasks complete. This feature is an evolution of earlier 'rough loop' concepts, but instead of a simple programmatic loop, it uses large language models to determine if a goal has been met, guiding the agent to keep working if it hasn't. This is particularly useful for ambiguous tasks, such as reducing Docker image size, where the exact steps aren't known upfront, allowing the agent to explore and iterate. To activate this in Code-X, users can enable the 'go' feature and then use the '/go' command followed by a detailed objective. Best practices for using this feature include defining what 'done' means very explicitly, providing clear constraints, validation methods, and stop rules, and engaging in an initial alignment conversation with the agent to provide context. Quantifiable stop conditions are crucial, rather than vague instructions. An open-source tool called 'go-body' can assist in constructing effective goal prompts. While this feature is effective for long coding sessions lasting hours, it is not yet designed for tasks spanning weeks or months that lack immediate verifiable results.