Architecture

Orchestrator agent with specialist agents

An executive reporting agent that understands the question and delegates to specialist agents for data, visualization, business context and email.

The problem

Leaders wanted to ask questions about incident and change data in plain English and get reports, charts and emails, not raw tables. One big prompt could not do all of that reliably.

The approach

• An orchestrator agent reads the user's intent and calls specialist agents as tools: data retrieval (text-to-SQL), visualization, business context, user context, feedback and email.
• Each specialist has a small, focused prompt and its own checks, so it can be tested and improved on its own.
• Model routing with a primary and a fallback model, and LLM-as-judge evaluation of answers.
• Co-designed with the team; I later rebuilt it as a LangGraph service with checkpointing, tracing (Langfuse) and logging (ELK).

What I learned

• Small specialist agents are easier to debug than one large agent.
• Tracing every step is what makes a multi-agent system maintainable in production.