Task planning is mandatory for complex missions such as wildfire fighting, extraterrestrial exploration, underwater discovery or military operations. These missions rely on the coordination of heterogeneous robotic fleets supervised by a single remote operator. The entanglement of operational constraints, agent heterogeneity, unforeseen events and intermittent communications can quickly exceed the operator's cognitive limits, leading to decision delays and degraded mission performance. First, we propose a comparison of the different application domains of automated planners, including aerial observation, space and underwater exploration as well as military operations. Although these missions appear very different, they share common challenges. This comparison allowed us to identify recurring issues and guide our various contributions. Then, our first contribution focuses on a shared-initiative architecture. The distribution of control between the autonomous system and the operator varies dynamically according to the operator's estimated cognitive load and the current state of the mission. The system continuously assesses the criticality of the situation and the operator's cognitive availability, then adapts its interaction strategy accordingly. This architecture aims to maximize the operator's understanding and apply their intent while protecting them from overload, including in the face of external events such as new task assignments, direct coactivity with an agent or unexpected environmental changes. Our second contribution is a planner designed to be integrated into our architecture. Its objective is to be expressive enough to encompass the complexity of the analyzed missions while remaining particularly efficient. It is based on constraint programming and enables the modeling of agent capabilities, temporal and spatial constraints, inter-agent synchronizations as well as the operator's cognitive limitations. This planner notably accounts for deployable agents and communication constraints with the command post. Performance results have validated its integration into the architecture. Finally, our third contribution is a replanning method designed to preserve plan stability when an unexpected event occurs. A distance metric between plans is defined to measure the structural gap between a new candidate plan and the plan previously communicated to the operator. This metric guides the solver's search toward structurally similar solutions, thereby reducing the operator's reappropriation effort and facilitating the understanding of the changes made. The system accompanies these modifications with explanations adapted to the level of detail required by the situation. The various contributions proposed in this thesis form a system and software components whose objective is to make the coordination of heterogeneous agents safer and more efficient in critical missions where a global overview is necessary to ensure the achievement of operational objectives. This work takes place in a context where autonomous robotic agents are increasingly common and where their proper use represents a major lever for operational efficiency gains.