Agent Integration R&D
A set of prototypes connecting LLM agents to tools, data and messaging channels. Under NDA, so this covers the architecture only.
Architecture
- Agent memory over MCP
- An MCP server and HTTP wrapper expose a Redis store to agents through a strict JSON contract.
- Messaging webhooks
- Webhook endpoints validate signatures and keep one conversation session per sender.
- Concurrent scraping pipeline
- Pluggable scrapers, one class per source, run concurrently with a cap, retries and exponential back-off.
- Streaming clients
- Real-time Server-Sent Events clients for voice-agent sessions.
- Backend proxy for agent flows
- Front ends talk to hosted agent flows through a backend proxy, never directly.
- LLM tracing
- Self-hosted tracing for model calls.
Main processing path
- channel webhooksignature check
- backend proxysession per sender
- agent flowLLM
- toolsMCP, data APIs
- replysame channel
Simplified, with generic component names. Highlighted steps use a model; the rest is software.