In this paper, we propose a flexible agent model, and a tool called AGENTφ for modeling safe composite agents. Agents are assembled from fine-grained reusable components, called micro-components, which implement non-functional mechanisms (interaction capabilities, mobility, adaptation skills...). Agents are self-adaptive : when running, an agent can autonomously replace some of its micro-components in order to adapt to runtime conditions. Here, the micro-component level behaves like an adaptable container which hosts the agent's behavior. Static modeling of agents and dynamic self-adaptation features allow developers to deal with flexibility requirements in the context of open distributed computing, and specially of ubiquitous and pervasive computing. An embedded communicating agent (with a GPS component) is developed in order to illustrate our proposal. Experiments show that complexity involved by flexibility constraints is limited due to separation of concerns and levels, and to reuse, still with good performances