FlowPlan: Reducing the Token Cost of Coding Agents with Abstraction-Guided Planning and Verification

Researchers introduce FlowPlan, an abstraction-guided repair approach for LLM-based coding agents that structures localization, planning, and verification around per-function control-flow or data-flow graphs. By reasoning at an abstract program level rather than repeatedly operating over raw source code, FlowPlan reduces exploration and post-edit verification overhead. Across benchmark settings, FlowPlan reduces token consumption while maintaining resolution rates relative to vanilla coding agents. Open article →