Joseph O'Sullivan Building a life-long learning agent The goal of building a life-long learning agent is to create an agent that, over a long lifetime in a single domain, can amortize effort expended in learning knowledge by using that knowledge to reduce the number of examples required to learn novel tasks. It has been assumed that, for instance, Multi-Task learning techniques are an appropriate basis for a lifelong learning agent. As such, I've constructed an MTL based lifelong learning agent and have applied it to a number of different synthetic testbeds -- classifying sonar data from Xavier's simulator, visual data from the ALVINN simulator, and performing function approximation. I use results from experiments in these domains to address 3 questions in this talk - what sort of benefit can be expected by the transfer of task-specific internal representations, how is the capacity and performance of the agent impacted as the number of tasks increases, and to what degree does ordering task presentation into a curriculum aid such a lifelong learning agent.