A new arXiv preprint introduces ADIAS — Automated Design of Interactive Agentic Systems — as a framework for building self-evolving agents. The core claim is that an agent can improve without updating its model weights, by iteratively optimizing the 'harness' — the surrounding interactive system that governs how the model is used — based on evaluation feedback.

The abstract contrasts this with existing methods, which it calls 'candidate-centric': they record evolution outcomes (the sentence is cut off in the source). ADIAS appears to shift the focus from candidate outputs to the design of the interactive system itself.

The source is a single arXiv abstract, and the provided text is truncated. No experimental results, benchmarks, or comparisons are available in the source, so the claims should be read as the authors' framing rather than verified findings.