Rill’s Agentic Architecture: Analytics for the AI Era

Rill's AI system evolved from using one-shot prompts to a sophisticated multi-agent architecture to enhance analytical workflows. Initially, Rill introduced a dashboard generator feature in February 2024, but faced challenges with scalability and maintenance due to disparate prompts. Inspired by Anthropic's AI agent strategies, Rill adopted a layered agent model where specialized agents, such as Developer and Analyst agents, share a unified runtime and tools while interacting with the same metrics layer as human users. This architecture ensures seamless collaboration between AI agents and humans by maintaining a single semantic layer that supports multiple clients without duplicating logic. The Developer agent focuses on project state management using a declarative approach and tool-driven workflows, while the Analyst agent interrogates the metrics layer to interpret data, ensuring grounded and verifiable analytics. Tools are treated as public APIs with schema-validated inputs and permission-aware access, enhancing reliability and consistency across internal and external agents. To maintain agent skill sync and optimize performance, Rill employs context engineering, pre-warming context, and bounding iterations, while also emphasizing evaluation with golden completions to fine-tune agents based on user feedback. This strategy minimizes hallucinations, enforces citation requirements, and ensures an efficient and accurate AI-driven analytics environment.