Multi-agent AI platform.
A production AI platform coordinating 18 specialized agents behind one user-facing interface.
What can be shared
This is an intentionally anonymized multi-agent system running in production. The public case focuses on the architecture and Isfan's engineering responsibility while omitting the product, company, industry, users, and operational workflows.
A single interface coordinates specialized agents and renders structured interactive responses. The underlying business actions and data sources are deliberately not disclosed.
The architecture decision
The first design was the obvious one: a single capable agent with many tools. In production it was too slow and hallucinated too often — one prompt carrying every responsibility degrades all of them.
The redesign split the work across 18 narrow agents: a triage agent in front classifies intent and hands the query to exactly one of seventeen specialists — product search, orders, consultation, and others — each with a small prompt and only the tools its job needs. Latency dropped and hallucination frequency fell, because every agent reasons over a fraction of the original context.
Integration
The agent interface is embedded in a Flutter host through a WebView + JS bridge, so the web layer can iterate quickly while the host remains native. Public details stop at that integration boundary.
The routing topology
Stack
Want the unabridged version — ask the agent or ask Isfan.