That's the question behind our paper accepted at ECIS 2026 in Milan: "Multi Agent Systems in the Lean Startup Cycle: Operationalising Dynamic Capabilities," with Elias Jelinek and Hannes Rothe.

We treat the Build–Measure–Learn loop as a closed-loop control system. Drawing on the Dynamic Capabilities View, we derive meta-requirements and design principles for a multi-agent system, then instantiate the artefact inside a production-grade SaaS codebase and test it against manual cycles in controlled simulations.

The result: time-to-validated-learning drops by roughly an order of magnitude — while preserving statistical rigour, full traceability, and nuanced Persevere/Iterate decisions. Sensing, seizing, and reconfiguring become observable at the feature level, and institutional memory accumulates in the artefact rather than only in the founders' heads.

In short: "agentic AI" shifts from a generic assistant to a disciplined experimentation infrastructure.