arXiv: AgentFactory — Self-Evolving Framework Stores Subagents as Executable Python Code, Not Textual Memory

arXiv·medium signalXBlueskyLinkedInCopy link

AgentFactory proposes a self-evolution paradigm where successful task solutions are preserved as executable Python subagent code rather than natural-language experience logs, enabling direct reuse and feedback-driven iterative refinement without manual intervention. Unlike prior approaches that summarize experience textually (lossy, un-executable, non-portable), AgentFactory subagents carry standardized documentation and work across any Python-capable system. The open-sourced framework demonstrates progressive capability accumulation — the subagent library grows and improves as more tasks are encountered, with concrete examples showing hardcoded implementations replaced by more robust solutions through iterative refinement.