On the stateless part - I increasingly believe that state keeping is an absolute necessity. Not necessarily across requests but on the local storage. Handoffs are proving invaluable in overcoming context limitations and I would like more tools to support a higher level of coordination and orchestration across sessions and with sub-agents.

I believe the best “worker” agents of the future are going to be great at following instructions, have a fantastic intuition but not so much knowledge. They’ll be very fast but will need to retain their learnings so they can build on it, rather than relearning everything in every request - which is slow and a complete waste a resources. Much like what Claude is trying to achieve with skills.

I’m not suggesting that every tool reinvent this paradigm in its own unique way. Perhaps we a single system that can do all the necessary state keeping so each tool can focus on doing its job really well.

Unfortunately, this is more art than science - for example, asking each model to carry out handoff in the expected way will be a challenge. Especially on current gen small models. But many people are using frontier models, that are slowly converging in their intuition and ability to comprehend instructions. So it might still be worth the effort.