The problem of performance maximization under a thermal constraint has been tackled by means of dynamic voltage and frequency scaling (DVFS) in many system-level optimization techniques. State-of-the-art ones have exploited Su-pervised Learning (SL) to develop models that predict power and performance characteristics of applications and temperature of the cores. Such predictions enable proactive and efficient optimization decisions that exploit performance potentials under a temperature constraint. SL- based models are built at design time based on training data generated considering specific environment settings, i.e., processor architecture, cooling system, ambient temperature, etc. Hence, these models cannot adapt at runtime to different environment settings. In contrast, Reinforcement Learning (RL) employs an agent that explores and learns the environment at runtime, and hence can adapt to its potential changes. Nonetheless, using an RL agent to perform optimization on manycores is challenging because of the inherent large state/action spaces that might hinder the agent's ability to converge. To get the advantages of RL while tackling this challenge, we employ for the first time multi -agent RL to perform thermally-restricted performance optimization for manycores through DVFS. We investigated two RL algorithms-Table-based Q-Learning (TQL) and Deep Q-Learning (DQL)-and demonstrated that the latter outperforms the former. Compared to the state of the art, our DQL delivers a significant performance improvement of 34.96% on average, while also guaranteeing thermally -safe operation on the manycore. Our evaluation reveals the runtime adaptability of our DQL to varying workloads and ambient temperatures.