# Governed Evolution of Agent Runtimes through Executable Operational Cognition

**Authors:** Mariano Garralda-Barrio  
**arXiv:** [2605.27328](https://arxiv.org/abs/2605.27328)  
**Date:** 2026-05-26  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2605.27328v1/

## Key points

- Recent advances in agentic systems have shifted the perspective on code to treat it as an executable operational substrate rather than a disposable output artifact.
- The framework introduced focuses on governed evolution in multi-agent systems through executable operational cognition, formalizing agent-generated artifacts as persistent runtime capabilities integrated into the operational substrate.
- The concept of HarnessMutation is key, enabling lifecycle-aware runtime adaptation under explicit validation, traceability, evaluation, and rollback constraints.
- Evolution is reframed as a bounded and observable process over persistent operational memory, emphasizing managed mutation control, lifecycle-aware selection, observability, graph-grounded representation, and governed reuse of artifacts.
- The proposed framework aims to optimize system behavior through agent-initiated artifacts that extend beyond individual tasks and ensures explicit, auditable, and constrained evolution in adaptive infrastructures powered by executable operational cognition.

## Summary

, , , , 

Recent advances in agentic systems have shifted the perspective on code, treating it as an executable operational substrate rather than a disposable output artifact. Building on prior work such as \emph{Code as Agent Harness}, which emphasizes the importance of validated agent-generated artifacts within runtime entities, this paper introduces a framework for governed evolution in multi-agent systems through executable operational cognition. In this framework, agent-generated artifacts are formalized as persistent runtime capabilities that gradually become integrated into the operational substrate rather than being transient outputs. A key concept introduced is \emph{HarnessMutation}, a mechanism for lifecycle-aware runtime adaptation that operates under explicit validation, traceability, evaluation, and rollback constraints. This approach reframes evolution as a bounded and observable process over persistent operational memory, moving away from unrestricted self-modification. The proposed framework not only provides a conceptual foundation for adaptive infrastructures but also offers practical insights on how these ideas can be implemented in modern agent runtimes and governance-oriented orchestration systems. By focusing on managed mutation control, lifecycle-aware selection, observability, graph-grounded representation, and governed reuse of artifacts, this framework aims to optimize system behavior through agent-initiated artifacts that extend beyond individual tasks. Overall, this refined summary highlights the shift from viewing code as task output to code as a lifecycle-managed capability within agentic systems. It positions the proposed framework as a crucial step towards ensuring explicit, auditable, and constrained evolution in adaptive infrastructures powered by executable operational cognition.