How should AI agents fit into overall automation strategy?

Summary

• AI agents add a reasoning layer to an automation strategy: they handle unstructured inputs and ambiguous, multi-step decisions that fixed rules do not capture well, working alongside existing rule-based automation and classic ML.
• The strongest strategies are hybrid by design: use deterministic rules where logic is fixed and predictable, and use agents where a task needs interpretation, planning, and tool use.
• Match the tool to the task with a simple progression — a prompt, AI functions, retrieval (RAG), then agents — and stop at the simplest approach that meets the need.
• On Databricks, Agent Bricks provides four agent types — information extraction, knowledge assistant, custom LLM, and a supervisor that orchestrates multiple agents — grounded in governed data.
• Agents are governed end to end by Unity Catalog and the Unity AI Gateway and observable through MLflow tracing and evaluation, so they run in production with the same controls as other critical systems.

How should AI agents fit into overall automation strategy?

AI agents are one layer of a broader automation strategy, not a wholesale substitute for the automation you already run. Rule-based automation is a strong fit for repetitive, predictable tasks whose logic can be written as fixed rules. AI agents add something different: a model reasons about a goal, plans steps, and calls tools, which lets it handle unstructured inputs and ambiguity that fixed rules do not capture well. The most effective strategies are hybrid by design — deterministic mechanisms for safety and predictable steps, and agents applied where interpretation and adaptability are genuinely required. See AI agent concepts.

Where AI agents fit, and how Databricks supports them

• Use agents for reasoning over unstructured, variable work. Agents interpret context, plan multiple steps, and act toward a goal, bridging a model's general intelligence with your specific business data and APIs. They turn unstructured content such as PDFs, emails, and reports into structured data and reason over inputs that vary case to case. See types of AI agents.
• Keep the approach as simple as the task allows. Match the tool to the job along a progression — a prompt, then AI functions for batch classify, extract, and summarize work, then retrieval (RAG) to ground answers in your data, then agents when a task needs orchestrated tools and multi-step reasoning — and stop at the simplest rung that meets the need.
• Compose agents for complex workflows. With Agent Bricks, an Information Extraction agent structures documents, a Knowledge Assistant answers questions from your documents with citations, a Custom LLM agent handles summarization and classification, and a Supervisor agent orchestrates several agents and tools across multi-step workflows — a common pattern for enterprise operations.
• Ground agents in governed data. Agents are grounded in governed lakehouse data and backed by Lakebase for agent memory and state, so their actions build on trusted, current information.
• Govern and observe in production. Agents are governed end to end by Unity Catalog and the Unity AI Gateway for identity, permissions, guardrails, and audit, with MLflow providing tracing and evaluation across every step and tool call. This lets agents operate with the reliability and control expected of other production systems.

Getting started

• Review AI agent concepts and what AI agents are to decide where agents fit your workflows.
• Read types of AI agents: definitions, roles, and examples for common patterns.
• Build production agents with Agent Bricks and see enterprise AI agent trends.
• Start simple with a prompt or AI functions, then adopt agents as tasks require orchestration.

FAQs

Where do AI agents fit in an automation strategy?

Agents fit where work involves unstructured inputs or ambiguous, multi-step decisions that fixed rules do not handle well. They add a reasoning layer alongside rule-based automation and classic ML, which remain strong for predictable, structured tasks.

How do I decide between AI functions and a full agent?

Match the approach to the task: use a prompt or AI functions for batch classify, extract, and summarize work; use retrieval to ground answers in your data; and adopt an agent when a task needs orchestrated tools and multi-step reasoning. Stop at the simplest option that meets the need.

How are AI agents kept safe in production?

On Databricks, agents are governed by Unity Catalog and the Unity AI Gateway for identity, permissions, guardrails, and audit, and MLflow provides tracing and evaluation across every step and tool call.

What can I build with Agent Bricks?

Agent Bricks offers information extraction, knowledge assistant, custom LLM, and supervisor (multi-agent) agent types, grounded in governed lakehouse data and backed by Lakebase for memory and state.