What is agentic machine learning?

Summary

• Agentic machine learning refers to AI-driven systems that can autonomously perceive, decide, and act to achieve a goal, rather than only producing output when prompted.
• Agentic systems have agency: they reason about how to proceed, execute plans, and call multiple tools or delegate to sub-agents to get work done.
• On Databricks, agentic ML bridges general intelligence and data intelligence — combining the AI model's capabilities with the specialized knowledge and APIs specific to your business.
• Genie Code (Public Preview) brings agentic development to ML workflows, helping teams build, debug, and improve production data science and ML systems from a prompt.
• The full lifecycle stays governed and observable, using MLflow tracing, Unity Catalog, and the Agent Framework to build, run, govern, and observe agents.

What is agentic machine learning?

Agentic machine learning refers to AI-driven systems that can autonomously perceive, decide, and act to achieve a goal. Rather than only producing output when prompted, an agentic system has agency: it can reason about how to proceed, execute plans, and call multiple tools or delegate to sub-agents. On Databricks, agentic ML bridges general intelligence, meaning the AI model's capabilities, with data intelligence, meaning the specialized knowledge and APIs specific to your business, so agents understand your data in context. See Agents concepts.

How Databricks supports agentic machine learning

• Agentic development for ML workflows. Genie Code (Public Preview) is Databricks' agent for data and ML work, helping teams build, debug, and improve production systems. Its Data Science Agent (Beta) can build an entire notebook for tasks such as exploratory data analysis, forecasting, and machine learning from a single prompt: it plans a solution, retrieves relevant assets, generates and runs code, uses cell outputs to improve results, fixes errors, and samples and visualizes data. See What's new with Genie Code at Data + AI Summit.
• Grounded in your ML stack. Genie Code integrates natively with the Databricks ML stack. It reads MLflow experimentation and observability data such as runs, artifacts, model lineage, and quality metrics; it inspects Model Serving endpoint health to diagnose serving issues; it moves to the appropriate compute when a job needs a GPU for training; and it writes features in your team's patterns. It learns from context that captures how your team builds features, trains models, and evaluates candidates, so it follows your patterns.
• Full ML lifecycle. An agent can explore data and create Delta tables, train and compare models, register models with MLflow and Unity Catalog, deploy a champion model to a serving endpoint, and diagnose and optimize running endpoints.
• Built, run, governed, and observed. The Agent Framework provides an end-to-end platform to build, run, govern, and observe AI agents. Agent Bricks can build a domain-specific agent from a task description and auto-generate evaluations, MLflow provides agent observability and tracing, and Unity Gateway acts as a control plane for rate limiting, guardrails, payload logging, and cost monitoring. See Mosaic AI announcements at Data + AI Summit.
• Evaluated and auditable. Agent development follows a build, evaluate, iterate, and deploy loop. Agents emit MLflow traces that record inputs, outputs, and metadata for every intermediate step, enabling systematic evaluation with LLM judges and human review, and providing an audit trail.

Getting started

1. Review the Agents concepts to understand agency, tools, and the agent lifecycle on Databricks.
2. Explore AI-assisted development and operations for how AI assists the ML lifecycle.
3. Try Genie Code (Public Preview) for data science and ML tasks, and use the Agent Framework to build, evaluate, and deploy agents with MLflow tracing and Unity Catalog governance.

FAQs

How is agentic ML different from a standalone AI model?

A standalone model produces output when prompted. An agentic system adds agency: it reasons about how to proceed, plans, and calls tools or delegates to sub-agents to complete a multi-step task autonomously.

What can an ML agent do across the model lifecycle?

It can prepare data and create Delta tables, train and compare candidate models, register models with MLflow and Unity Catalog, deploy a champion model to a serving endpoint, and help diagnose and optimize endpoints in production.

How are agentic ML systems kept trustworthy?

Agents emit MLflow traces capturing every step, which supports evaluation with LLM judges and human review and provides an audit trail, while Unity Catalog and Unity Gateway provide governance and control.