This paper proposes a multi-agent system architecture based on introspective agents provided with an expressive communication language and reasoning capabilities on their skills. The agents' introspective capabilities make it possible to agents to autonomously decompose a task according to their competencies. Using a protocol, the agents can then coordinate with other agents to as to collectively and dynamically cover the performance of a complex task with interrelationships and dependence among its actions. We describe in this paper the introspective agents' structure and the coordination protocol that supports the distributed and decentralized task achievement.