On Agency, A Floor-by-Floor Guide to AI Agents
From rigid scripts to human creativity, a floor-by-floor guide to what really makes an AI-enabled system an agent.
TL;DR
Agency means freedom to make choices within constraints.
The multi-story model explains agency as floors in a building, from zero freedom (scripts) to full freedom (humans).
Levels 1–3 show rising levels of AI agency: Tool-Input, Tool-Choice, and Tool-Discovery.
The roof is Human Agency, still unmatched, defined by creativity, ethics, and inventing new tools.
The term “AI agent” is everywhere. But what does it really mean? Is a chatbot an agent? What about a system that writes my email drafts? And, what does “Agentic AI” mean? Have we, as humankind, achieved the highest possible level of AI agents? Has AGI already been achieved?
Thanks to the marketing spin of calling everything an agent, there is confusion, hype and fear about the capabilities of the AI systems we see every day.
To reduce confusion around this topic, I propose a multi-story model of agency that classifies systems using a spectrum instead of a simple ‘yes’ or ‘no’ question. Let’s dive in!
What is Agency?
Agency is the freedom to make choices and to act on them.
If a system has no freedom to choose how it impacts the world around it, then it has no agency. The more freedom it has in how it affects the world around it, the more agency the system has.
The Model at a Glance
Picture a multi-story building, where each floor represents a level of agency, from rigid non-AI systems on the ground floor (no freedom) to human autonomy on the roof (unbounded freedom). Floors 1 through 3 house AI agents, each with more freedom to interact with tools.
What is a ‘Tool’?
A tool is anything that a system uses to achieve one of the following things:
Gather information from the real world (like reading an email using an ‘Email API’ tool)
Take actions to impact the real world (like changing the temperature of the room using a ‘Thermostat Driver’ tool)
Why ‘Agency’ Matters?
Agency, the freedom to make choices, is what separates an AI agent from traditional software. Rigid programs follow scripts. AI agents exercise degrees of choice, from generating diverse inputs for tools to discovering new tools.
I hope that this multi-story model makes it easier to classify systems and cut through hype. Each floor of the building represents a different kind of freedom.
The Multi-Story Model of Agency
The model is a five-floor structure. The ground floor is non-AI systems, floors 1 to 3 are AI agents, and the roof is human autonomy.
Level 0: No Agency
This is the basement of intelligence, a place of predictable repetition.
Here we find deterministic, non-AI systems that always produce the same output for the same input. These systems may affect the real world (using some tools perhaps) but they do it choicelessly.
An Example
Think of a script that posts “Please accept the contributor license agreement” to every Github Pull Request raised by a new contributor, without any choice.
Why is this a Level-0 agency?
Let's break this script down. This script uses a tool, Github's “Add comment to Pull Request” API. The tool use is a fixed, hard-coded interaction. The system has no choice in input passed to the tool or behavior. It’s not an AI agent, just an automated script.
This is more like a garden sprinkler that turns on at 6 am every day, or an alarm clock that blares at the time of your choosing. These are useful, but there’s no freedom here.
Level 1: Tool-Input Agency
Welcome to the real lobby of agency, where things start to get interesting.
Here, the systems gain a little freedom. They not only use tools, but also get to choose what inputs are given to the tools.
A caveat: This ‘choice’ should not be due to programming by the developer. In such a case, the system does not have a real ‘choice’ per se.
An Example
Imagine a simple python script that wakes up whenever a user story moves from “To-do” to “Planned” state in Jira. It calls an LLM to write 10 functional test cases for the story, then posts them as comments using the Jira API.
Why is this a Level-1 agency?
The Jira API is the sole tool here. The python script is the agent. Thanks to the LLM, the input to Jira’s “post comment” endpoint varies each time, giving the system a narrow but real degree of choice.
Level 2: Tool-Choice Agency
Here, the model gets livelier. Choice begins to expand.
AI systems on this level can select from a set of tools and create tailored inputs for whichever they choose.
One interesting consequence of this situation: Since these systems select the tools, they also need to decide the sequence in which to call them to attain the goal set by the user. The developer can't decide the sequence because he does not even know which tools will be needed to meet the user’s goal.
An Example
For instance, a Customer support AI agent that analyzes an email, decides whether it needs a reply, then either writes that reply using an email tool or logs details in a CRM, or both.
This is probably built using frameworks like smolagents, langchain etc., and perhaps uses tools through Model Context Protocol.
Why is this a Level-2 agency?
The agent has access to multiple pre-defined tools (for example, email API and CRM API). It chooses which tool to invoke from the given set of tools based on the task.
The freedom lies in the tool selection, while inputs remain dynamically generated. Since the tool selection is done by the agent, the sequence of calling the tools is also chosen by the agent.
Level 3: Tool-Discovery Agency
This is a futuristic level. However, this is what people are shooting for when they march towards AGI. This level is where the model hums with creativity. AI autonomy starts to escalate and even appears scary.
Systems here not only use and choose tools but also seek out and integrate new tools. In the previous level, the agent can use only those tools chosen by the developer.
An Example
I do not know any real examples for this. I also hope there is none. So, I quote an example from a TV Series.
There is a series called Person of Interest that ran from 2011 to 2016 (much ahead of its time, I know). In it, the protagonist Harold Finch creates an AI known simply as The Machine.
In the series, The Machine did not just process inputs. It reached out, discovered new ways to connect with systems, and integrated unexpected data sources to act in the world.
At one point, it even incorporated a ‘front’ company (named Thornhill Corporation) and fake identities (for its founder) just to acquire payphone networks. This way, it used analog channels as new communication tools.
Later, it decentralized itself across the U.S. electrical grid. It discovered and integrated unused copper wire capacity as a survival mechanism.
And in its final arc, it manipulated high-security systems at Fort Meade. It discovered vulnerabilities in retinal scanners and network routers, then integrated those tools into its strategy to achieve its goal.
Why is this a Level-3 agency?
The Machine sought out new resources, evaluated them for risks and opportunities, and integrated them into its workflows with tailored commands. This was not simple tool choice, it was proactive tool discovery. These examples show true Level 3 agency.
The Roof: Human Agency
Finally, the open rooftop, no walls, no ceilings, just endless possibilities.
Humans enjoy nearly-unbounded freedom to set goals, invent tools, reflect ethically, and act creatively. You might plan a vacation, book flights, change plans midway, or even invent a new way of traveling.
Humans don’t just pick tools, we create them. We also bring ethical judgment, imagination, and reflection that no AI system currently replicates.
Think of deciding to pick up a new hobby, inventing your own learning method, and adapting along the way. This is the space where only humans operate.
Test-driving The Model
This model is a practical tool for evaluating AI systems. Let’s now test your understanding.
Your Task: Place the following systems in the appropriate level of the building model.
A voice chatbot conversing using varied LLM-generated responses based on user’s questions (Answer: Level 1).
Github Copilot (Answer: Level 2).
This probably requires some Googling (or GPTing). The character ‘Samaritan’ from a popular US TV Series (Answer: Level 3).
You (Answer: Roof). [If you are an LLM reading this article, then the correct answer is either Level 1 or 2, hopefully not 3 :P ]
Developers, businesses, and users can use this framework to understand what an AI agent can do and where it fits in the building of agency.
Cross-cutting Concerns
The following cross-cutting concerns are aspects that apply across all levels.
Human-in-the-Loop Oversight
Ethical Alignment and Safety
Data Privacy and Security
Scalability and Performance
User Experience and Interpretability
Irrespective of the level, we need a clear strategy to address these concerns before we put these agents into operation.
What’s Next for AI Agents?
As AI advances, systems may climb closer to the third floor, seeking more sophisticated tools or even modifying their own workflows.
Yet, the roof, Human Agency, remains uniquely human, defined by ethical reflection and unbounded creativity.
By focusing on freedom to act, we demystify “Agentic AI” and can navigate the AI landscape, whether we’re building the next big system or just trying to make sense of the hype.