How developers are using AI agent plans to guide autonomous coding tools

Developers working on open-source projects are increasingly creating structured Agent Plans—task-specific guidance files that help autonomous AI coding tools navigate complex software engineering work. A new exploratory study from researchers examines how these artifacts function within real-world repositories and what kinds of instructions they contain.

The study scope#

The research team examined 36,710 GitHub repositories from engineered software projects, screening them for evidence of Agent Plan files stored in tool-specific directories. They ultimately identified 85 Markdown plan files spread across just 10 repositories—a concentrated but revealing sample. The findings, detailed in a preprint, suggest that while Agent Plans remain a niche artifact, they offer meaningful insights into human-agent collaboration.

What Agent Plans support#

The identified plans guided AI agents across a diverse range of development activities:

• Maintenance and bug fixes within existing codebases
• Design decisions for new features or architectural choices
• Code construction and implementation work
• Quality assurance, testing, and validation tasks
• Process support for workflow management

How they guide agent execution#

Rather than being generic instructions, Agent Plans tend to be highly task-oriented. They most often include implementation steps that break down what the AI should do, specific file locations and paths so the agent knows where to focus, and testing and validation information to verify correctness.

This kind of repository-preserved guidance appears to represent an emerging best practice for teams that deploy autonomous coding systems—a way to encode task intent and execution strategy at the project level.