Agent should learn to alternate between performing long research and record the findings.
This is because event stream summarization might destroy context, making the research cut off. Therefore, the agent should "cache" its own research for a long task/research sequence.