The generative AI (GenAI) revolution has fundamentally changed how organizations extract value from data. Large language models (LLMs) excel at understanding and generating human-like text, but their true enterprise value emerges only when they can access proprietary data and take real-world action.

While vector databases and retrieval-augmented generation (RAG) gave LLMs memory, Amazon Nova provides execution and specialization.

In this article, we break down the Amazon Nova model family, with a deep focus on Nova Act and Nova Forge, and explain how they enable a shift from passive chatbots to autonomous, enterprise-grade AI agents.

What Is the Amazon Nova Model Family?

Amazon Nova is not a single model, but a tiered family of foundation models designed to balance intelligence, latency, and cost.

*Note: Model pricing may differ based on region

Instead of choosing between large, expensive models and smaller, underpowered ones, Nova provides a structured hierarchy:

• Nova Micro – Text-only model optimized for ultra-low latency and high-volume tasks
• Nova Lite – Multimodal model (text, image, video) with an optimal speed-cost balance
• Nova Pro – Advanced reasoning model for complex logic, coding, and math
• Nova Premier – High-capability “teacher” model used for advanced reasoning and synthetic data generation

While these models form the foundation, the real engineering innovation lies in how AWS enables them to act and specialize through Nova Act and Nova Forge.

Amazon Nova Act: Enabling Autonomous AI Agents

Most LLMs operate in a text-in, text-out loop. Ask them to “help me navigate this local government website to apply for a passport,” and they can explain the steps, but they cannot execute them.

Amazon Nova Act changes this paradigm.

Nova Act is an agentic workflow service that allows Nova models to interact directly with web-based systems using a secure, AWS-managed headless browser environment.

How Nova Act Works

Unlike brittle automation tools such as Selenium, which often break when a website’s structure changes. Nova Act relies on visual and semantic understanding, not static selectors.

The workflow follows four stages:

1. See (Visual Parsing)
   The model captures both the Document Object Model (DOM) and the rendered visual state of the page.
2. Plan (Intent Analysis)
   A natural-language instruction (for example, “Find the January invoice”) is translated into a structured sequence of steps.
3. Act (Interaction Execution)
   The agent performs actions such as clicking, scrolling, and typing within a secure browser container.
4. Verify (Outcome Validation)
   The system confirms whether the expected state change occurred (for example, verifying that a file download has started).

Why Nova Act Matters

This capability unlocks Actionable AI. These are AI systems that don’t just retrieve information but execute real workflows, including:

• Navigating supplier portals to collect pricing data
• Automating end-to-end QA testing for web applications
• Downloading, parsing, and analyzing financial or compliance documents

In effect, Nova Act turns AI into a digital worker, not just a conversational interface.

Nova Act Playground

To have a hands-on-experience on how Nova Act works, you can try their playgrounds to get a grasp on how you will set up agents.

Amazon Nova Forge: Building Custom, Domain-Specific Intelligence

Foundation models are powerful, but they are generalists. In enterprise environments, generic intelligence is rarely enough.

Amazon Nova Forge is a fine-tuning and distillation environment designed to create smaller, faster, and highly specialized models trained on your proprietary data.

The Limitations of Generic LLMs

A general-purpose LLM may understand poetry, trivia, and casual conversation, but enterprise systems often require models that deeply understand:

• Proprietary data schemas (e.g., internal JSON logs)
• Industry-specific terminology (e.g., medical or legal codes)
• Organization-specific workflows and policies

The Solution: Distillation and Data Blending

Nova Forge introduces two critical techniques:

1. Model Distillation

A high-capability teacher model (Nova Premier) generates high-quality responses, often using synthetic data, which are then used to train a smaller student model (Nova Lite or Nova Micro).

Result:
You retain much of the teacher model’s intelligence while achieving significantly lower latency and cost.

2. Data Blending

During training, proprietary enterprise data is blended with Amazon’s curated general-reasoning datasets.

Result:
This prevents catastrophic forgetting, ensuring the model remains fluent, coherent, and capable of logical reasoning, even after specialization.

Choosing the Right Nova Component: A Practical Comparison

Conclusion

Just as vector databases became the memory layer for modern AI systems, Amazon Nova is positioning itself as the action layer.

By combining:

• Autonomous agents (Nova Act)
• Custom model distillation (Nova Forge)
• A scalable family of foundation models

AWS enables developers to build AI systems that do more than generate text–they execute, adapt, and scale.

Getting Started with Amazon Nova

You can explore Amazon Nova models and agent capabilities with or without Amazon Bedrock. Here are some of the resources you could explore:

• Amazon Nova Chat
• Amazon Nova User Guide 
• Building Agents with Amazon Bedrock
• Amazon Nova Act Documentation
• Amazon Nova Act Playgrounds
• Amazon Nova Forge Documentation