by2olak | Zoho

600 users. Paper contracts. Handwritten Arabic and English documents stacked in filing cabinets. The CRM was there, but nobody could get it to do what the business actually needed. The system had to understand Zoho from the inside out. Not just read records, but discover modules, understand relationships between them, and act on them without being told exactly what to do.

A master orchestrator built in Python with Claude Agent SDK sits at the top. It coordinates a fleet of specialized sub-agents, each responsible for specific Zoho CRM modules through LangChain Deep Agents. Google's A2A protocol handles inter-agent communication, so agents can delegate work to each other without routing everything through the orchestrator. The master discovers available Zoho modules dynamically at runtime. New Zoho functionality becomes available to the system without touching the orchestration code.

CopilotKit provides the interface through generative UI. The system does not present a fixed dashboard. It builds the interface you need based on what you are trying to do. Ask it to move a contact through a pipeline stage and it materializes the right form, pre-filled with context pulled directly from the CRM. Role-based agents queue tasks from persistent state, so nothing falls through the cracks when a user closes the browser.

An OCR agent processes handwritten Arabic and English documents autonomously, extracting structured data from paper contracts and intake forms before they ever touch the CRM. Everything runs on-premise. Zero data leaves the building. Human-in-the-loop validation checkpoints ensure accuracy before any CRM updates are committed. A stepping stone to SAP.