Iterative Program Refinement With Failure Driven Learning

via “dynamic code refinement through error-driven iteration”

Agent that uses executable code as actions.

Unique: Closes the error-recovery loop by feeding execution errors back to the LLM with full context, enabling agents to self-correct code iteratively. Tracks refinement history and enforces iteration limits.

vs others: More autonomous than systems requiring human intervention for error fixes, but slower than systems that avoid errors through careful prompt engineering