We propose a new real-time load sharing policy (LSP), which optimally dispatches the incoming workload according to the current availability of the operators. Optimality means here that the global service permanently requires the engagement of a minimum number of operators while still respecting due dates. To cope with inherent randomness due to operator failures as well as non-stationary fluctuating incoming workload, any optimal LSP rule will necessarily rely on real-time updating mechanisms. Accordingly, a permanent monitoring of the traffic workload, of the queue contents and of other relevant dynamic state variables is often realized by a central workload dispatcher. In this contribution, we abandon such a "classical" approach and we propose a fully decentralized algorithm which fulfils the optimal load sharing process. The underlying decentralized decisions rely on a "smart tasks" paradigm in which each incoming task is endowed with an autonomous routing decision mechanism. Incoming jobs hence possess, in this paper, the status of autonomous agents endowed with "local intelligence". Stigmergic interactions between these agents cause the optimal LSP to emerge. We emphasize that beside a manifest strict relevance for applications, our class of models is analytically tractable, a rather uncommon feature when dealing with multi-agent dynamics and complex adaptive logistics systems.