How Mosaic AI Enables the Next Generation of Agentic Systems

Mosaic AI Overview

What sets Mosaic AI apart is how it brings together three critical pieces: data, models and governance into one framework. Many organisations have good data lakes, or large language models, or smart analytics but struggle to unite them, manage scale or govern them. Mosaic AI makes this easier.

Unified data & intelligence: Structured and unstructured data live in the same platform, meaning you can search through knowledge bases, run vector similarity on embeddings and pull features in real time.

Governed intelligence: With things like Unity Catalog, you don’t just build agents quietly instead you track every decision, every tool call, every model version.

AI native architecture: From fine tuning models to indexing vectors to orchestrating multi agent systems, Mosaic AI is built for the demands of modern generative and agentic AI.

Continuous learning: The platform supports feedback loops, evaluation and monitoring, so your agent doesn’t stagnate, it gets better.

How Mosaic AI Works

1. Mosaic AI Agent Framework & Agent Bricks

The core of building intelligent agents on Databricks lies in its dedicated tooling designed high quality.

Mosaic AI agent framework: This is the modular, open-source compatible SDK for crafting complex AI agent systems. It provides the necessary plumbing for Retrieval Augmented Generation (RAG), managing tool calling logic and coordinating multi agent systems using frameworks like LangChain, LangGraph and CrewAI. Critically, it ensures that your agent code and configuration are logged as traceable MLflow Models for governance and deployment.

Agent bricks (low/no code): If you want to get results quickly without diving deep into code, Agent Bricks makes things straightforward. It gives teams an easy to use interface where they can build, test and fine-tune different types of agents like ones that help with knowledge management or pull out specific information from data without needing to write much (or any) code. Plus, it can generate practice data and automatically adjust settings to make sure everything works well, this includes capabilities for synthetic data generation and automated tuning to achieve high quality.

2. Unifying Governance with Unity Catalog (UC)

When it comes to enterprise AI, governance isn't optional it's essential. Unity Catalog (UC) helps you maintain control not just over your regular data tables, but across your entire AI agent workflow from start to finish.

End-to-end access control: UC governs access to all assets an agent touches:

Data and vector indexes: Fine grained security (row/column level) applies directly to the underlying data and the Mosaic AI Vector Search indexes that serve as the agent’s long-term memory.

Models and functions: It governs permissions for the served LLMs and the Unity Catalog Functions that act as agent tools.

Agent assets: The agent's code, configuration and execution logs are managed and versioned via UC, ensuring that only authorised agents can access and execute specific business logic.

Auditability and lineage: UC provides detailed audit logs for agent actions, tracing the data used, the model called and the decision made. This is vital for compliance and debugging, ensuring you can answer the “why” behind any autonomous action.

Certified assets: New features like Catalog tags (e.g., ‘certified’, ‘deprecated’) help teams discover and use trusted, compliant models and tools, preventing agents from operating on unverified resources.

3. Tool Orchestration: Tools, Genie Space and MCP

Agentic AI systems are defined by their ability to use tools. Databricks provides a governed, extensible mechanism for this:

Unity Catalog functions (tools): Any SQL or Python logic from running a complex calculation to querying a secure Delta table can be registered as a Unity Catalog Function. The agent treats these as internal, governed tools, ensuring that tool execution honours the user's UC permissions.

Genie space: This refers to the AI/BI Genie capability, which allows agents to use natural language to query and interact with structured data in the Lakehouse. Agents can be given access to a Genie Space as a powerful, specialised tool for generating analytics, charts and data insights.

Model context protocol (MCP): This new protocol and its associated MCP Catalog in Databricks Marketplace are key for extending agents to external, trusted services (e.g., specialised web indexes or proprietary APIs).This means your agents can safely work with external tools and services, while Databricks governance still keeps everything in check through AI Gateway and Unity Catalog Connections. You get full visibility into what's happening with outside systems and everything stays compliant with your policies.

4. MLOps, Testing and Observability with MLflow 3.0

The old MLOps approaches don't really fit in when you're working with agents. That's why MLflow 3.0 was specifically designed to handle the unique challenges of managing generative AI throughout its entire lifecycle.

Tracing and observability: MLflow Tracing lets you see exactly what's happening when your agent runs. It records everything that went in, what came out, the prompts being used and all the tool calls or LLM interactions happening along the way. This becomes important when you need to figure out why something went wrong in a complicated, multi-step reasoning process.

Generative AI evaluation: MLflow 3.0 doesn't just rely on basic performance numbers. Instead, it offers:

• LLM judges: These automatically evaluate more subjective aspects like how relevant or safe the outputs are. You can even fine tune these judges using plain language feedback so they match your specific industry requirements.

• Review app: This gives your subject matter experts an easy to use interface where they can provide organised feedback. That feedback then becomes the foundation for building solid evaluation datasets.

Deployment jobs: MLflow 3.0 has added deployment jobs (which Unity Catalog governs) to handle the entire testing and rollout process automatically. Basically, it makes sure no agent version goes live until it's passed all quality checkpoints.

5. AI Gateway: Centralised Model Access and Guardrails

The Mosaic AI Gateway acts as the central traffic control layer, ensuring secure, consistent and cost-effective access to all generative models.

Unified model access: It gives you one centrally managed API endpoint where you can access any model you need, whether that's commercial ones like OpenAI GPT-5 or Anthropic Claude Sonnet 4, open source options like Llama, or your own custom models running on Databricks. This means your teams can swap between different models without having to rewrite any of their agent code.

Security and compliance: The Gateway makes sure important AI safety measures and security rules are in place:

• PII filtering: It catches and blocks Personally Identifiable Information (PII) as it happens, both in what goes into the system and what comes out.

• Safety filtering: It screens content for harmful material like hate speech and violent content.

Monitoring and cost control: It gives you a central place to manage rate limits, control who has access to what and track how everything's being used (all of this gets logged to Delta tables in Unity Catalog). This way, administrators can keep a close eye on both the costs and operations of all LLM usage happening across the organisation.

Key Advantages of Mosaic AI

Unified data: Mosaic AI brings data, models and governance together in one place.

Governed intelligence: Unity Catalog makes sure access stays secure, tracks where data comes from and how it's used and manages all your policies in one place.

Scalable multi-agent orchestration: Easy set up and manage AI agents that work together across different areas of your business.

Automated deployment workflow: Everything you need, like model serving, RAG components and evaluation tools, is already built in, so it becomes easier to use.

Feedback driven growth: Through MLflow tracking, agents evolve with real world feedback and usage patterns.

Challenges & Considerations

Latency: Complex multi agent orchestration can increase response time; optimisation is key.

Accuracy: RAG pipelines reduce hallucination risk but require high quality embeddings.

Cost management: Embedding and inference costs must be balanced against business value.

Conclusion

Mosaic AI is a revolutionary enterprise intelligence platform that combines structured analytics with generative AI, enabling organisations to move beyond static reports to living systems of intelligence that continuously learn, reason and adapt. It can be used for improving retail operations, optimising supply chains and transforming financial risk analysis, enabling smarter, faster and more contextual decision-making.

With Cloudaeon’s expertise in AI orchestration, governance and enterprise scale deployment, organisations can now move beyond traditional analytics and embrace AI systems that think, converse and create tangible business value.

Interested in knowing what Databricks Mosaic AI can do for your organisation? Talk to an expert now!