What is a custom AI agent and when should I build one?

Summary

• A custom AI agent is an AI system that perceives, decides, and acts to reach a goal — using a model as the "brain" that reasons, calls tools or sub-agents, and returns an answer or asks for clarification.
• Databricks recommends a default-to-managed approach: start with managed Agent Bricks templates and build a custom agent only after exhausting their configuration.
• Build custom when you need deterministic coded control flow, a mandated framework, tools a managed framework cannot host, persistent memory across sessions, or custom UI and business logic — or when managed quality is still insufficient after tuning.
• For most use cases — document Q&A, text-to-SQL analytics, multi-tool reasoning — managed Agent Bricks (Knowledge Assistant, Multi-Agent Supervisor, Genie) covers the need.
• Even custom agents stay governed: wrap any framework in an MLflow ResponsesAgent, deploy on Databricks Apps, and keep Unity Catalog governance, tracing, and evaluation.

What is a custom AI agent and when should I build one?

An AI agent is a system that can autonomously perceive, decide, and act to achieve a goal. A model acts as the "brain": it receives a request, reasons about how to proceed (which data to fetch, which tools to call, which logic to apply), executes a plan by calling tools or delegating to sub-agents, and returns an answer or asks for clarification. A custom AI agent is one you build with your own control flow, tools, and logic when a managed agent cannot meet the requirement. On Databricks, the recommended path is to default to managed Agent Bricks agents and build custom only when you need to. See introducing Agent Bricks and agent concepts.

Why Databricks Agent Bricks for custom AI agents

• Default to managed, build custom when needed. Databricks recommends starting with managed Agent Bricks templates and building a custom agent only after exhausting their configuration knobs. Managed templates cover common patterns such as document Q&A, text-to-SQL analytics, and multi-tool reasoning. See building trusted AI agents: choose, govern, and scale with confidence.
• When to build a custom agent. Consider a custom agent when you need deterministic coded control flow (exact branch and loop sequences rather than model routing), a standardized framework (for example LangGraph or CrewAI), tools the managed framework cannot host, persistent memory across steps and sessions, or custom UI and business logic to embed the agent in a larger application — or when quality remains insufficient after tuning a managed agent.
• Managed templates for common needs. Knowledge Assistant chunks and vectorizes documents, selects embedding and LLM models, tunes retrieval, and deploys production-ready APIs with source citations; Multi-Agent Supervisor routes across Genie (text-to-SQL), Knowledge Assistant, Unity Catalog functions, MCP servers, and AI Search; and AI functions (ai_query, ai_parse_document, ai_extract) turn documents into structured data in SQL pipelines. See what is agentic analytics.
• Custom agents stay governed. When managed templates do not suffice, wrap any framework in an MLflow ResponsesAgent and deploy it as a Databricks App — the agent stays traced, evaluated, and governed on Databricks, with Unity Catalog access control and lineage and the AI Gateway for governed model access. See shipping quality enterprise AI agents with Agent Bricks and Databricks Apps.
• One governed path from build to production. Agent Bricks and Databricks Apps provide a single path to build, evaluate, deploy with SSO and fine-grained permissions, and continuously improve agents on your own governed data. See introducing Agent Bricks.

Getting started

• Read introducing Agent Bricks.
• Learn what an AI agent is and how it works.
• See building trusted AI agents: choose, govern, and scale with confidence.
• Ship an agent to business users with Agent Bricks and Databricks Apps.

FAQs

What is a custom AI agent?

A custom AI agent is an agent you build with your own control flow, tools, memory, and business logic — using a model to reason and act — for cases a managed agent template cannot meet.

When should I build a custom agent rather than use a managed one?

When you need deterministic coded control flow, a specific framework, tools the managed option cannot host, persistent memory across sessions, or custom UI and logic — or when managed quality remains insufficient after tuning.

What can managed Agent Bricks agents handle without custom code?

Common patterns such as document Q&A, text-to-SQL analytics, and multi-tool reasoning, using Knowledge Assistant, Multi-Agent Supervisor, and Genie.

Do custom agents lose governance?

No. Wrapped in an MLflow ResponsesAgent and deployed on Databricks Apps, a custom agent stays governed by Unity Catalog and remains traced and evaluated.