Choosing among Agent Bricks (Knowledge Assistant, Multi-agent Supervisor, Information Extraction) for a problem

Agent Bricks provides pre-built, auto-optimized agent templates—Knowledge Assistant, Multi-Agent Supervisor, and Information Extraction—so engineers can assemble production-quality GenAI apps without hand-tuning prompts or writing custom evaluation harnesses. Choosing the right Agent Bricks type depends on whether the task is grounded Q&A, orchestration across multiple specialized agents, or structured data extraction from unstructured content.

1 · Learn the must-know

• 1Agent Bricks auto-optimizes each agent's quality, cost, and latency using built-in LLM-judge evaluation against a task description and optional labeled examples, minimizing manual prompt engineering.
• 2Knowledge Assistant is designed for RAG-style question answering grounded in a curated unstructured corpus (e.g., documents in Unity Catalog volumes) and automatically handles chunking, retrieval, and grounding.
• 3Information Extraction is the right choice when the goal is converting unstructured text or documents into structured fields/JSON output at scale, not for open-ended conversation.
• 4Multi-Agent Supervisor is used when a single agent can't cover the needed scope—it routes requests and orchestrates multiple specialized agents, tools, or Genie spaces to answer complex, multi-domain queries.
• 5Pick Agent Bricks over a fully custom agent build when you want a managed, evaluation-driven path to production; drop to the custom Mosaic AI Agent Framework when requirements exceed what a template supports.
• 6All Agent Bricks agents are deployed via Model Serving, governed through Unity Catalog, and integrate with MLflow for tracing and ongoing evaluation.

2 · Check your understanding