Nathan Lambert Defines the AI Agent Spectrum

What is an AI Agent?

AI agents are digital tools that interact with the world, both online and offline. Unlike chatbots, which mimic human conversation, AI agents aim to perform specific tasks. However, the term “AI agent” has been too broadly defined, often conflating simple autonomy with complex decision-making tasks. Nathan Lambert, a machine learning researcher with a PhD from Berkeley, provides a breakdown of where AI agents are headed and what needs to be defined more clearly for these tools to truly benefit us going forward.

“The current definition of AI agent that most people use encompasses way too much under one term. This goes back to the obsession with all agents being viewed through the lens of reinforcement learning. We need taxonomies for different types of agents. As I started working more towards directions for 2025 related to agents, as in planning, basic tool use, long-context, etc., an obvious mismatch appeared between the discourse and reality of agents.

The most basic AI agents that people are going to be using in 2025 are things like language models with search, which actively pull from knowledge stores, and something like Siri in Apple Intelligence, which acts as an orchestration layer on top of the operating system.1 The simplest way to view the starting points for language model-based agents is any tool-use language model. The spectrum of agents increases in complexity from here.

In the current zeitgeist, an “AI agent” is anything that interacts with the digital or physical world during its output token stream.”

Nathan Lambert, The AI Agent Spectrum

Types of AI Agents

The spectrum ranges from basic language models aiding searches, to more advanced systems like Siri that manage various software tasks. Moving beyond this, agents will increasingly operate autonomously, blurring traditional computing boundaries.

Here are four terms that Lambert has defined to classify and understand the role of a specific AI agent:

• Scoping the Horizon: How long can the agent run for and how does it plan across time?
• Defining Utility: How is the AI agent succeeding or not?
• Pruning Information: How is the agent interfacing with other digital information, both on inputs and outputs?
• Interacting with Agents: Is the agent designed to interact with other agents?

Current AI agents span from simple tools to more orchestrative models capable of integrating multiple operations. Existing examples include email assistants with varying capabilities—from basic email sorting to comprehensive digital management.

The Impact of AI Agents

Agents will not only perform individual tasks but complex, interconnected processes. Understanding how to design and instruct these agents effectively will become crucial for productivity. This doesn’t require coding skills but rather the ability to deconstruct tasks and provide precise instructions. As these agents grow in capability, they will create interconnected networks automating routine tasks and freeing humans from mundane work.

Questions and Challenges Ahead with Agents

Several questions loom as we move towards this integrated AI future. Will these models learn and adapt through online training? What happens when AI agents start interacting with each other, potentially creating an entirely new online ecosystem? How do we best manage and regulate these agents? And finally, how will interacting with such agents feel for end users?

This technology promises to reshape everyday digital interactions, pushing AI from a tool into a system fundamentally altering how we work and live. For deeper insights, Nathan Lambert’s Substack, (https://www.interconnects.ai/), is a great resource.