AgentForge uses role-play to teach novices agentic software engineering
Practice as Planner, Reviewer, or Tester while AI agents handle the rest
Agentic AI systems are increasingly automating software development tasks, but they often operate as black boxes, making it hard for novices to understand or validate AI decisions. AgentForge, from researchers Zihan Fang, Yueke Zhang, and Yu Huang, tackles this by letting learners adopt a specific software-engineering role while AI agents handle the other three positions in a multi-agent code-repair workflow. This role-based scaffolding exposes how agents coordinate, share artifacts, and make decisions, turning abstract AI behavior into a hands-on learning experience.
In a study with 37 novice developers, AgentForge produced high task-completion rates, but the learning curve varied sharply by role. The Code Reviewer practice demanded significantly more interaction turns, reroutes, and completion time, and was rated the most difficult. Yet across all roles, participants reported substantial gains in their understanding of software repair and agent collaboration (p_adj < .001). These results suggest that immersive role-play can help novices build practical software-engineering skills while learning to critically guide and evaluate agentic AI output—a crucial competency as AI-assisted development becomes standard.
- AgentForge assigns novices to one of four roles (Planner, Patch Author, Code Reviewer, Test Runner) in a multi-agent code-repair workflow
- Study of 37 novice developers showed high task completion, but Code Reviewer required significantly more interaction turns and time (p_adj = .004)
- Participants reported strong learning gains in software repair and agent collaboration (p_adj < .001) from the role-based scaffolding
Why It Matters
As AI agents join dev teams, AgentForge offers a practical, scalable way to train humans in critical oversight.