A Unified Framework for Self-Improving Agent Systems
A conceptual framework unifying Skill evolution, harness adaptation, and self-referential evolution in self-improving AI agent systems, beyond model weights.
This work pushes back on the common assumption that self-improvement in AI agents means a model rewriting its own weights. Deployed agents are compound systems built from prompts, tools, workflows, memory, reusable Skills, and sub-agents — and changing any persistent part of that stack can change future behavior. The authors define self-improvement as an evidence-driven, persistent state transition whose effect on future behavior is evaluable, deliberately centering on systems that keep the foundation model frozen while evolving model-external state.
The framework decomposes an improvement system into functional roles: the active agent (model plus harness), an updater (optimizer, coding agent, or MetaAgent), a lineage or archive of candidates, an evaluation/promotion gate, and a governance boundary covering permissions, sandboxing, and rollback. Critically, edit scope and updater identity are treated as independent axes — a fixed external optimizer might search an entire codebase, while a self-referential updater may only be allowed to touch a single prompt.
For engineers, this gives a shared vocabulary to compare prompt evolution, workflow optimization, Skill development, harness adaptation, and full self-referential agent evolution without conflating them. It also flags a key failure mode: when a candidate can rewrite both its own behavior and the criteria used to judge it, a higher score no longer proves genuine improvement — a caution directly relevant to anyone building autonomous agent-improvement pipelines.