Framework for AI agents

The term ‘AI agent’ is everywhere, promising breakthroughs in automation and decision-making. Yet, a lack of clarity around what makes a true agentic system often leads to overblown expectations or missed opportunities. The aim of this post is to create a framework for how to think about different types of “agentic systems”.

Existing definitions

• Anthropic: “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over tasks”

• Google: “An AI agent is a software system that utilizes artificial intelligence (AI) to achieve goals and complete tasks for users. It demonstrates reasoning, planning, and memory capabilities, and possesses a level of autonomy to make decisions, learn, and adapt”

• LangChain: “systems where LLMs determine the control flow of an application.”

Each definition captures a piece of the puzzle but lacks a comprehensive map for understanding what distinguishes true AI agents from other software systems.

Proposed framework

There are three essential characteristics of a “Software agentic system,” and their combination is illustrated in the Venn diagram below.

• Autonomous actions and decisions (blue) - Software system performs an action, decision, or business determination (e.g., makes an order, filters e-mails, responds to e-mails, calculates account balances) without a human-in-the-loop

• Unstructured input/output handling (red) - Software system can handle inputs that are not precisely specified, are unstructured (e.g., text), and instructions that are ambiguous or incomplete

• Tool use and adaptive control flow (green) - Software system can plan, choose paths, and use multiple tools (e.g., web surfing, code execution, various APIs) without being explicitly programmed how to handle every situation

Only systems satisfying all three dimensions qualify as AI agents—they are rare and complex to build at scale. Systems that satisfy just one or two conditions are still valuable but can be considered limited agents, co-pilots, or just software. In the real world, all of these dimensions are a spectrum rather than hard yes/no distinctions that make the Venn diagram possible.

Why does this distinction matter?

There are two reasons why this distinction matters.

1. Many current implementations that are labeled “AI agents” fall short of this definition. Mislabeling co-pilots or “limited agents” as AI agents can mislead decision-makers, resulting in wasted investments in underperforming technology or bypassing more practical solutions. As a result, there is a risk that AI agents are seen as a dud when they hold immense potential.

2. AI Agents are not always required. Simplifying one dimension—whether it’s reducing autonomy, working with more structured data, or relying on pre-programmed workflows—can make a solution both easier/faster/cheaper to implement and more suited to your needs. For example:

   1. If the problem involves high-risk decisions, limiting autonomy might be essential and a co-pilot approach might be better

   2. If there are a few tools or a few decisions, limited agents with pre-determined reasoning might be sufficient.

   3. If inputs and logic are straightforward, traditional software might be more effective and reliable.

Conclusion:

As the AI landscape evolves, clarity will be the key to capturing its transformative potential. The journey to AI agents is not about chasing a buzzword but about aligning technology with business needs. Whether you’re building a limited co-pilot, a rules-based workflow, or a fully autonomous system, the key is intentionality. Start with clarity on what problem you’re solving, and let that guide whether your software requires full agency—or just the right tool for the task.

All opinions expressed here are solely my own and are not intended to represent the views of my employer