MOSS: Self-evolving autonomous agent systems via source-level rewriting

Researchers have developed methods enabling autonomous agents to modify and improve their own source code without human intervention, treating code generation as a continuous self-refinement loop rather than a one-shot task.

This addresses a core constraint in deployed agentic systems: agents currently require human-in-the-loop updates to improve performance or adapt to new domains. Self-rewriting agents compress the feedback cycle and reduce operational dependency on engineering teams for iterative agent tuning. For long-horizon tasks, this enables agents to accumulate learned patterns in executable form rather than only in weights or memory.

Operators deploying such systems face new governance questions: monitoring agent code drift becomes essential, as self-modifications may accumulate in unpredictable ways. The cost of manual agent oversight decreases, but observability infrastructure becomes non-negotiable. Teams will need version control and rollback mechanisms for agent codebases, treating autonomous code generation as a production system rather than a development artifact.