AI-Native Software Ecosystems
Autonomous coding agents now write, review, and merge code alongside each other. Each one can pass its own tests while the codebase they share accumulates problems that belong to no one. This research line builds the theory, and the evidence, for software that is no longer the sum of its contributions.
The wrong unit of analysis
The field evaluates coding agents the way it has always evaluated components: one at a time, on isolated benchmark tasks. That worked while AI stayed a tool in a developer’s hand. It stops working when agents become contributors, because contributors interact, and interaction produces effects no per-agent benchmark can see. The question that organises this line is simple to state and hard to answer: when a shared codebase degrades, does the problem belong to any individual agent, or to the ecosystem where it accumulates?
More is different
The theory paper opens with Philip Anderson’s 1972 argument: knowing the fundamental laws of the small does not let you reconstruct the behaviour of the large. More Is Different applies that argument to software. AI-native ecosystems must be studied as complex adaptive systems, because their most consequential properties, architectural entropy, cascade failures, comprehension debt, are emergent: they arise between agents, not inside any one of them. The paper maps Holland’s six properties of complex adaptive systems onto observable ecosystem dynamics, defines the state variables and coarse-graining functions that make emergence measurable rather than metaphorical, and derives seven propositions that revisit Lehman’s laws of software evolution where their agent-level assumptions no longer hold. Every proposition states what would confirm it and what would refute it. Falsifiable is the point.
The risk lives in the repository.
Govern the Repository, Not the Agent puts the theory against data at scale: more than 930,000 agent-authored pull requests, measured through integration friction, the cost of landing a contribution in a codebase that others are changing at the same time. The result is uncomfortable for per-agent evaluation. About half of the variation in friction stays with the repository after the contribution, its author, its size, and its agent are all accounted for. And agent-authored contributions concentrate that repository-level friction roughly twice as strongly as human ones, an intraclass correlation of 0.30 against 0.16, a gap that survives controls for codebase size, age, task shape, process maturity, and merge path.
The evidence base
Measured, not asserted
arXiv 2604.19827 · 2606.28235
Govern the ecosystem, not the agent
If the risk is a property of the ecosystem, then certifying agents one at a time answers the wrong question, however rigorous the benchmark. The repository is where the risk collects, and it is the level at which the risk can be observed, measured, and managed. For research, that argues for ecosystem-level monitoring as the primary governance mechanism of AI-native software. For organisations adopting autonomous agents, the practical question is not which agent to trust, but how to watch the codebase they share.
Papers
Russo, D. (2026). More Is Different: Toward a Theory of Emergence in AI-Native Software Ecosystems. arXiv:2604.19827
Russo, D. (2026). Govern the Repository, Not the Agent: Measuring Ecosystem-Level Risk in AI-Native Software. arXiv:2606.28235
Where this research programme is heading.
This line extends the arc that runs through the AI in Software Engineering pillar: from how engineers adopt AI tools, through the AI4SE1DK national programme, to ecosystems where agents are contributors in their own right. Both papers are open access, and the line is open to research collaboration.