Loosely-Structured Software: Engineering Context, Structure, and Evolution Entropy in Runtime-Rewired Multi-Agent Systems

A team of researchers has published a seminal paper proposing a new class of software systems called Loosely-Structured Software (LSS). The work addresses a critical scaling problem: as LLM-based multi-agent systems (MAS) grow in size and autonomy, their free-form interactions create 'runtime entropy'—unpredictable complexity, coordination errors, and system drift. The authors argue that traditional software engineering, focused on deterministic logic, is insufficient for managing systems where agents can generate and rewrite their own code during execution.

To make this chaos governable, the paper introduces a three-layer engineering framework. First, View/Context Engineering manages the execution environment and task-relevant perspectives for agents. Second, Structure Engineering organizes the dynamic binding between software artifacts and the agents that manipulate them. Third, and most crucially, Evolution Engineering provides principles to govern the lifecycle of these self-rewriting artifacts. Building on this, the team proposes LSS design patterns, which act as semantic control blocks to stabilize interactions while preserving agent adaptability. The goal is to provide the architectural discipline needed to build robust, large-scale agentic infrastructure that can evolve reliably over time.