What are the best AI agent frameworks?

Summary

• An AI agent framework provides the building blocks to create, deploy, and operate AI agents: orchestration and multi-agent coordination, secure tool use, memory, evaluation, governance, and model serving.
• The best frameworks treat governance and evaluation as first-class, so agents run on governed data and tools and their quality is measured continuously, not bolted on later.
• On Databricks, Agent Bricks and the Mosaic AI Agent Framework provide an end-to-end path to build, evaluate, deploy, and govern agents.
• Tools are governed through Unity Catalog with support for the Model Context Protocol (MCP), on-behalf-of-user authorization, lineage, and audit trails, so agents respect the data permissions of the people who invoke them.
• Built-in evaluation with MLflow, one-line deployment to Mosaic AI Model Serving, Lakebase-backed memory, and Unity AI Gateway governance round out the framework.

What are the best AI agent frameworks?

An AI agent framework provides the software building blocks for systems that can reason over a task, call tools, remember context, and act. The strongest frameworks cover the full lifecycle, not just the first prototype: orchestration and multi-agent coordination, secure access to tools and data, memory for state, evaluation to measure quality and cost, governance and audit, and production serving. Because agents act on real data and systems, governance and evaluation matter as much as orchestration. On Databricks, Agent Bricks and the Mosaic AI Agent Framework deliver this end to end on a governed platform.

What to look for in an AI agent framework

• Orchestration. Multi-step reasoning, tool sequencing, and coordination across multiple cooperating agents.
• Secure tool use. Standardized, governed access to APIs, databases, and SaaS systems, ideally using the same permission model as your data.
• Memory. Persistent state and conversation history so agents stay context-aware across steps and sessions.
• Evaluation. Quality, cost, and latency measurement with tracing that shows what the agent did at each step.
• Governance. Access control, policy enforcement, and audit trails across models, tools, and data.
• Model serving. Production deployment with auto-scaling and the flexibility to route across multiple models.

Why Databricks Agent Bricks and the Mosaic AI Agent Framework

• Build, framework-agnostic. Agent Bricks generates domain-specific synthetic data, builds task-aware benchmarks, and auto-optimizes agents for cost and quality. The Mosaic AI Agent Framework covers build, test, deploy, and monitor, and supports authoring agents with popular open-source agent libraries as well as the Databricks Agent SDK. Agents run on serverless compute with built-in security and governance. See author an agent and custom agents.
• Governed tool use through Unity Catalog. Agent governance and data governance are defined in the same catalog. Agents access tools via managed MCP servers for Databricks data, external MCP servers registered as governed services, Unity Catalog functions for custom logic, or custom tools built with the Agent SDK. On-behalf-of-user authorization enforces that agents respect the data permissions of the person invoking them, complete lineage traces agent outputs back to source Delta tables, and every model and tool call is recorded in the Unity Catalog audit trail. See what is tool calling.
• Memory on the lakehouse. Lakebase provides built-in memory powered by the lakehouse, giving agents a production-grade way to store and reuse state and conversation history without extra infrastructure.
• Built-in evaluation with MLflow. Mosaic AI Agent Evaluation combines AI judges for accuracy, hallucination, and harmfulness with human feedback and root-cause tracing. MLflow Tracing records inputs, outputs, and metadata for every intermediate step, custom LLM judges score retrieval and response quality, and a built-in review app collects user feedback, with the same evaluation applied across development, staging, and production. See agent evaluation and monitoring.
• Model serving and model choice. Mosaic AI Model Serving provides serverless, auto-scaling endpoints for ML models, LLMs, and agents, with one-line deployment and batch or real-time serving. Native access to leading proprietary and open foundation models gives teams model optionality and cost flexibility.
• Governance with Unity AI Gateway. The Unity AI Gateway is a single governed control plane for the AI estate, adding access control and contextual policies, rate limiting, PII and safety guardrails, payload logging, and usage and cost tracking across every request and response.
• Multi-agent orchestration. Agent Bricks includes a Multi-Agent Supervisor for orchestrating cooperating agents into a single workflow.

Getting started

• Start with Agent Bricks and the governed enterprise agent platform overview.
• Author an agent with the Agent Framework and connect governed tools per tool calling.
• Evaluate with MLflow and deploy to Model Serving.

FAQs

What is an AI agent framework?

An AI agent framework is the set of building blocks for creating and operating AI agents: orchestration, tool use, memory, evaluation, governance, and serving. It lets teams move an agent from prototype to a governed, monitored production system.

How does Databricks govern the tools an agent can use?

Tools are governed through Unity Catalog, using the same permission model as data. Agents connect via MCP servers, Unity Catalog functions, or custom tools, and on-behalf-of-user authorization plus audit logging ensure agents respect the invoker's data permissions and every call is recorded.

How do you evaluate agents on Databricks?

Mosaic AI Agent Evaluation combines AI judges, human feedback, and MLflow Tracing that records every step, with custom LLM judges and a review app. The same evaluation configuration applies from development through production.

Can I use my own agent libraries?

Yes. The Mosaic AI Agent Framework is framework-agnostic and supports authoring agents with popular open-source agent libraries as well as the Databricks Agent SDK, then deploying them to governed, serverless model serving.