Self-Evolving Coding Agents
The paper frames coding agents as systems that can persistently improve from their own repair work.
The survey defines “self-evolving coding agents” and separates them from conventional coding agents. It maps what can change: framework, memory, skills, tools, model components, workflows, topology, environment, and context. It also tracks when those changes happen and what software-specific signals drive them, including executable feedback and repository context. The authors flag hard problems around unreliable feedback, benchmark overfitting, reversibility, complexity, safety, cost, and generalization. HF Daily Papers' note
The survey defines “self-evolving coding agents” and separates them from conventional coding agents. It maps what can change: framework, memory, skills, tools, model components, workflows, topology, environment, and context. It also tracks when those changes happen and what software-specific signals drive them, including executable feedback and repository context. The authors flag hard problems around unreliable feedback, benchmark overfitting, reversibility, complexity, safety, cost, and generalization. HF Daily Papers' note
score 4