A "Useful" Definition for AI Agents

Defining AI agents has been all the rage these last eight months. While we’re not particularly interested in esoteric definition debates, we eventually came to a practical and useful definition for AI agents that helps us in our day-to-day. At conferences, meetups, consulting, and in discussions with friends, many have found our definition clarifying. This post explains our “useful definition” position and we hope you find it useful as well.

Lately I’ve been enjoying more episodes from the a16z podcast, the official podcast from the venture capital firm Andreessen Horowitz. In a recent episode they reviewed the oft-debated question, “What is an AI Agent?”

Part of me agrees with one of the participants, Matt Bornstein, pointing out that straining over the definition of a made-up concept is a bit of a waste of time. He said, “I think we’ve all just been nerd-sniped. I just think it’s, like, humanities people love classifying and they draw kind of like fine distinctions between different types of things, entities, whatever.”

However, in our day-to-day work we settled on a clear, functional definition that has been helpful for coordinating work among teams. Instead of a buzzword to toss around, “agentic” and “agent” have practical definitions that help us align planning and actions across teams.

We settled on this definition only once we settled on an overarching framework of the underlying properties for what is and is not an agent.

A Definition and a Framework

Agents are components within systems where AIs can orchestrate their own processes and take AI actions to accomplish tasks.

For clarity, we break it down into a simple 2x2 matrix made up of two parts, orchestration and action.

Orchestration – Does the AI orchestrate the solution?

Action – Can AI be used to take an action?

That’s it. Apple’s Siri was an AI Router, using AI (primarily through NLP) to decide what to do. An AI workflow would be an LLM evaluating a refund request and making a recommendation based on company policy. An Agent does both the orchestration and can choose to take AI actions.

Below is a more detailed explanation, and we’ll provide more in-depth examples in follow-up posts (e.g., how “agentic” is a system?). But for today, we’ve found this a clean place to start from and enhance communication across the team.

Skip to Next Part Unless You Want the Gory Details

Briefly walking through each quadrant:

1. AI Orchestration + AI Actions = Agents

Full AI agents. An AI directs the processes and tools, actively determining how tasks are executed to achieve its goals. The AI agent decides which AI workflows and tools, along with their corresponding actions, to trigger.

Agents offer the greatest flexibility and adaptability. However, complexity, uncertainty, and risk also peak here.

2. AI Orchestration + Non-AI Actions = AI Routers

These systems orchestrate decisions, but rely on non-AI actions or tools. These tools are often predefined system processes. Virtual assistants such as Siri or Alexa typify this quadrant. They decide how to respond to user queries, yet the actual actions, such as playing a song, setting an alarm, or sending a message, remain relatively straightforward and predetermined.

AI Routers simplify execution. They retain adaptive benefits of AI orchestration, but with lower complexity and fewer uncertain outcomes due to the simpler actions executed.

3. Non-AI Orchestration + AI Actions = AI Workflows

This quadrant represents the traditional AI workflows where non-AI orchestration triggers and manages predefined workflows, and these workflows perform AI-driven actions.

AI Workflows allow your teams to harness powerful AI capabilities while controlling orchestration through predictable, rules-based methods. This reduces complexity and simplifies management, at the cost of adaptability.

4. Non-AI Orchestration + Non-AI Actions = Traditional Software Engineering (SWE)

Here we find traditional, rules-based software systems with no significant AI integration. An example would be an automated email-reporting system triggered each morning by a scheduled job. Both orchestration (cron jobs, scheduled triggers) and actions (sending a formatted report) follow predefined rules without AI-driven decisions or adaptive capabilities.

Traditional SWE offers maximum predictability, reliability, and simplicity, but lacks adaptability, innovation, and the powerful analytical leverage provided by AI.

No quadrant is inherently superior. The optimal choice depends on the objectives, the complexity of tasks, and the level of adaptability required.

Contrasting Definitions

How does our definition match up with many definitions offered by organizations, companies, academics, and your neighbor’s dog? Of all of the definitions we’ve seen bandied about, Anthropic’s definition aligns closest to what we feel like is an effective practical definition:

"Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."
– Engineering at Anthropic

This definition captures a concrete distinction. Workflows follow deterministic paths, which are a series of steps executed consistently each time. In contrast, AI agents select and orchestrate among available workflows and tools, continuously evaluating options and outcomes until they achieve their goal, determine it's not worthwhile to proceed, or are deactivated. However, we diverge from Anthropic's framing in one key aspect: AI orchestration does not necessarily require an LLM. Instead, orchestration might be executed through alternative forms of AI, potentially ones better optimized for rapid feedback loops and decision making.

In the end, don’t get too hung up on definitions. These are evolving terms. However, we’ve settled on a practical definition that lines up relatively closely with one of the major companies and enhances communication among team members. Among the technical members, this works. If you’re looking for how to craft marketing hype for AI agents… that’s for a different post.

You can catch a16z’s podcast episode on “What is an AI Agent?” here if you’re interested: https://a16z.com/podcast/what-is-an-ai-agent/

(I highly recommend the podcast, especially if you’re the type who knows which episodes early on you should skip.)