Started with a complex vision:
- Backend profiles for each agent
- Context presets with drag-and-drop
- Tool-generated context pieces
- Dynamic context compaction
But then we found clarity through simplification.
Just system prompts with:
- Name
- Content (the actual prompt)
- Backend (which LLM to use)
- CRUD operations
Each agent gets assigned one system prompt. Clean, simple, useful.
The exploration revealed interesting needs:
- Tools generating context pieces
- Dynamic context management
- Registered/unregistered pieces
But these are solutions looking for problems we haven't hit yet. The immediate need is just: different agents need different prompts and backends.
This is a familiar pattern in our work:
- Explore the full vision
- Find the minimal useful core
- Build that first
- Let real use drive complexity
The complex context system might be exactly what we need... later. For now, system prompts are enough.