Zoho Chief Scientist on Agent Architecture
Sridhar Vembu, Chief Scientist at Zoho, outlined a practical architecture for software agents. The model involves agents making API calls driven by LLMs, which are then passed through a validation layer, with human supervision providing the high-level objectives. He noted this moves beyond simple conversational user interfaces toward functional, supervised automation.
- Zoho's AI stack, which underpins its agent architecture, includes the proprietary Zia LLM, offered in multiple sizes (e.g., 1.3B, 2.6B, 7B parameters) to optimize for specific business tasks like RAG and code generation. The ecosystem is supported by a no-code Zia Agent Studio with over 700 pre-built actions and a Model Context Protocol (MCP) server designed to allow third-party agents to interact with Zoho's applications securely. - The architecture described is commonly implemented using open-source LLM orchestration frameworks. LangChain and its extension LangGraph are used for building stateful, multi-agent applications, while Microsoft's Agent Framework (combining AutoGen and Semantic Kernel) is geared toward enterprise-grade, multi-agent systems integrated with Azure. - In insurance, similar agentic systems are achieving measurable ROI; commercial P&C insurers have seen loss ratio improvements of 3-5% and reduced quote-to-bind times by 60-99%. Specific use cases include automating the entire claims lifecycle, from intake and triage to fraud detection, reducing operational costs by up to 30%. - Advanced implementations often rely on multi-agent design patterns to coordinate complex workflows. These include the orchestrator-worker pattern, where a central agent assigns tasks, and hierarchical patterns, where higher-level agents delegate to and supervise specialized lower-level agents. - Architecting the backend to support these agents requires an API-first, event-driven approach to enable real-time responsiveness. Scalable designs use asynchronous processing with task queues (like RabbitMQ or Kafka), containerization with Kubernetes for auto-scaling, and API gateways to manage authentication and rate-limiting. - For aspiring technical founders, the insurtech venture market has reset after a 2021 peak, with funding stabilizing around $4.25 billion in 2024. Investor focus has shifted dramatically to B2B SaaS solutions, which captured 43% of insurtech VC funding in 2024
