# Agent Orchestrator

*Origin: Frontend design skill test evolved into a real work project idea*

## Problem Statement

We have multiple AI agents at work but no unified interface:
- Natural language SQL search agents
- Company knowledge base agents
- [Likely others...]

Users currently bounce between different interfaces or don't know which agent to use for their query.

## Solution Concept

A single **Orchestrator** interface that:
- Provides one entry point for all agent queries
- Intelligently routes requests to the appropriate specialized agent
- Shows visual feedback of the routing decision
- Maintains conversation context across agent handoffs
- Allows manual agent override when needed

## Visual Direction

From the initial frontend test:
- Dark corporate aesthetic with warm gold accents
- Editorial typography (Crimson Pro + Plus Jakarta Sans)
- **Key feature**: Visual orchestration flow showing the handoff in real-time
- Sidebar with specialized agents (color-coded by function)
- Auto-resizing input with agent selector

## Current Agent Inventory

| Agent | Function | Status |
|-------|----------|--------|
| SQL Search | Natural language to SQL queries | Active |
| Knowledge Base | Company docs/knowledge search | Active |
| [Others?] | | |

## Architecture Thoughts

### Intent Classification
- Route to "sql" when query involves data/tables/metrics
- Route to "knowledge" when query asks about policies/procedures
- Fallback to general or multi-agent mode

### Context Persistence
- Conversation thread maintained across agent handoffs
- Results from one agent can inform another
- Shared memory for the full session

### Agent Registry
- Dynamic discovery of available agents
- Each agent defines its capabilities and routing triggers
- Easy to add new agents

## Open Questions

- [ ] What agents exist beyond SQL and knowledge?
- [ ] How do agents currently expose their APIs?
- [ ] Authentication/authorization model?
- [ ] Self-hosted vs SaaS deployment?
- [ ] Existing tech stack constraints?

## Next Steps

1. Inventory all existing agents and their interfaces
2. Define the orchestrator API contract
3. Determine routing logic (LLM-based vs rules-based)
4. Build POC with 2-3 agents
5. Design system architecture diagram

## Related Files

- Initial UI prototype: `frontend/subagent-orchestrator-chat/index.html`

---

*Elaborate further here...*