Building a Multi-Agent System from 23 Open-Source Projects — What Worked, What Broke

The Single-Agent Ceiling One LLM, one context window, one shot. That is the constraint every AI agent operates under today. You can give an agent tokens - 200K, 1M, even 2M context windows exist now - but that only pushes the ceiling higher, it does not remove it. A single agent still has to hold everything in one head: research, planning, coding, debugging, deployment. When context runs out, it forgets what it knew three turns ago. The obvious solution is multiple agents that hand off work to each other.