The phenomenon of construction delays is a well-known cause of cost overruns, and the problem is usually triggered by something as simple as the illness of a group of workers, bad weather, or any other reason why the plan must deviate from the predetermined schedule. Learned planners solve this problem by coming up with an entirely new plan and rarely do they promise that the learned action takes precedence into account. This work frames daily repair as a multi-agent decision problem on a single 79-task residential construction instance, where a knowledge graph is compiled into a structural feasibility mask that gates every action so that whatever schedule the policy produces is feasible by construction, not by hope. A shared-parameter policy trained with Centralized Training with Decentralized Execution and Proximal Policy Optimization (CTDE-PPO) under heavy disturbances recovers a mean of 20.48 days relative to a static-replay baseline (240.84 ± 9.05 vs. 220.36 ± 7.50 days), with zero constraint violations and instability an order of magnitude below a reactive heuristic; the 95% confidence interval on recovery, [11.2, 29.8] days, excludes zero. Disabling the mask, by contrast, produces thousands of violations almost immediately. A Constraint Programming-Boolean Satisfiability (CP-SAT) solver builds a stronger static plan than either baseline, yet the rescheduler still closes much of that gap online, through nothing more than daily adaptation. Together these results point to a workable path for bringing guaranteed-feasible Multi-Agent Reinforcement Learning (MARL) into operational construction management.

multi-agent reinforcement learning, knowledge graphs, construction scheduling, feasibility masking, proximal policy optimization, project rescheduling