Multidisciplinary engineering problems are commonly solved by coupling software components that are developed for each discipline independently. This has motivated the optimization community to develop optimization architectures that support modularity. The individual discipline feasible (IDF) formulation achieves this goal by introducing new optimization variables and nonlinear constraints that effectively decouple the discipline solutions during each optimization iteration. Unfortunately, the added number of variables and constraints can be significant. Conventional optimization algorithms may exhibit poor convergence rates due to large numbers of design variables, and the explicit constraint Jacobian required by most is prohibitively expensive to compute for IDF constraints. To address these challenges, we propose a reduced-space inexact-Newton-Krylov algorithm that leverages matrix-free KKT matrix-vector products via second-order adjoints, a GMRES-based Krylov iterative method tailored for non-convex problems, and an effective matrix-free preconditioner for the IDF problem. We implement this algorithm using a parallel-agnostic reverse communication interface, and demonstrate its efficacy on a variety of PDE-constrained optimization problems.