I’ve been working for some time now with a team inside of Microsoft on various ideas in the direction of more stable, longer-running agents based on LLMs, that use memory and other techniques keep context and perform more complex tasks, or even just maintain relationships with people (which is pretty wild).