These are self hosted for learning experience, I could have built an agent swarm in the cloud, but I'd never have learnt the fundamentals.

- Cold starts impact, context length issues, task lifecycle management

- Inefficiencies in delegation, necessitating workflow patterns for small projects (big AIs hide this problem until you scale and they hit the same issues).

- Limits of the AI would be harder to find or notice (e.g. where time - and cost - is being spent needlessly).

Or just use a more capable local model like Qwen3.8-Flash-Next?