Production Agentic AI System for Monite API

I developed an advanced multi-agent AI system for the financial platform at Monite, which enabled users to interact with financial data in Monite API using natural language. The project solved critical problems of traditional agentic approaches: unpredictable behavior and hallucinations when working with complex financial data.

From “Magic Monolithic Box” to Engineering Solution

The initial single-agent approach showed critical shortcomings:

• Rapid context overflow after 2-3 dialogues
• Hallucinations with financial data
• Non-deterministic behavior, effectively unacceptable for fintech
• Tool-call chaos with dozens of functions

The architectural solution was an orchestrated multi-agent pipeline with explicit distribution of responsibilities for each sub-agent.

The project uses the SoTA Schema Guided Reasoning pattern instead of regular prompting, and the semantic-router library approach for deterministic determination of internal functions that extract data to serve one or more interests from user queries (interests/intents are extracted from queries in advance).

SGR directs the attention of each agent to systematic “reasoning” through a preset schema.

The target system became predictable, transparent for tracking results and debugging, and most importantly - excellently open to functional extension. The architecture allows adding dozens and hundreds of functions without serious trade-offs in accuracy.