N-version programming is a practice, discussed in Agentic Software Engineering: Foundational Pillars and a Research Roadmap, of generating multiple independent versions of a solution to the same task. The paper describes its re-emergence in Agentic Software Engineering: a team of agents can generate several candidate pull requests for the same ticket in parallel, which serves both as a form of inference-time compute that increases the probability of a successful outcome through trial and error, and as a means of creative exploration.
Usage
In the paper’s motivating example, a developer resolving seven tickets triggers an agent team to generate 28 candidate pull requests in parallel (four per ticket); the developer then selects the most promising candidate per ticket, or refines the specification and re-triggers the agents if none is acceptable. The paper states its proposed Agent Command Environment (ACE) (ACE) must support “disciplined N-version programming,” letting a developer visualize, compare, and mix components from multiple agent-generated solutions. A LoopScript’s task-decomposition-and-parallelization capability — assigning one BriefingScript to multiple agents — is what the paper says makes N-version programming “routine,” shifting the key metric from single-task latency to overall system throughput. It also notes N-version programming as a case where a coach’s guidance may involve synthesizing a final solution from different agent-generated drafts (e.g. combining the UI from one solution with the backend logic from another).
When It Applies
The paper treats N-version programming as an existing, decades-old fault-tolerance technique being repurposed for agentic coding rather than one it originates; it cites the practice’s re-emergence as made practical by the ability to run multiple LLM-based agents (or heterogeneous teams of specialized agents, such as one model for high-level planning and another for detailed code generation) in parallel against the same specification.