The use of distributed computation and control is pervasive for a wide range of smart grid research topics. However, recent developments of smart grid co-simulation platforms have not been able to provide effective support for the modeling and simulation of distributed software systems. In particular, cosimulation literatures only focused on the integration of electrical and communication network simulators, and the responsibility for modeling distributed software is often delegated to one of these two simulators. Since these domain specific simulators are not designed for this purpose, such delegation incurs many limitations that prevent convenient, effective, and accurate modeling of software behaviors. To mitigate the problem, this paper presents, to our knowledge, the first co-simulation integration of direct-execution simulators to provide dedicated support for distributed smart grid software. We first present the development of the novel DecompositionJ framework (DEterministic, COncurrent Multi-PrOcessing SImulaTION for Java programs), which is a compiler-based code analyzer and transformer to automatically convert multi-thread Java programs into direct-execution simulators, eliminating the need for manual code or model development. Next, we apply DecompositionJ framework to generate simulators for a popular multi-agent platform JADE. The JADE simulators are then integrated with electromagnetic transient simulator (PSCAD) and packet-level network simulator (OPNET) using standardized co-simulation runtime infrastructure. At last, we conduct a case study on agent-based smart grid restoration using this novel co-simulation platform. Through the analysis of simulation results, it is shown that the proposed direct-execution simulation framework is able to facilitate the understanding, evaluation, and debugging of distributed smart grid software.