Article
Jul 12, 2026
Agentic AI: When Assistance Becomes Action
Imagine asking for directions.
Now imagine not having to ask at all.
The system already noticed you paused at the intersection. It already checked the crosswalk signal. It already decided what you needed to know, and when.
That shift, from responding to acting, is what people mean when they talk about agentic AI.
From Answering to Acting
Most AI systems people have used so far are reactive.
You ask a question. It gives an answer.
You issue a command. It executes that command.
The relationship is transactional. Input, then output. Nothing happens without a prompt.
Agentic AI works differently.
An agentic system doesn't just answer questions. It pursues goals. It breaks a broad objective into smaller steps, decides which step comes next, takes action, observes what happened, and adjusts.
It doesn't wait to be asked at every turn.
This is a meaningful departure. And it raises a harder question than "what can this system do?"
The harder question is: how much should it decide on its own?
Theoretical Foundations
Fig. 1 - The perceive-plan-act loop. The agent never stops cycling through it; each action reshapes the environment it will perceive next.
It perceives the current state of the world. It plans, deciding what action would move it closer to its goal. It acts, and that action changes the environment, which the agent then perceives again.
This loop is old. It describes a thermostat almost as well as it describes a research robot. What separates a thermostat from an agentic AI system is the complexity of the "plan" stage, how many steps ahead it reasons, how many goals it weighs at once, and whether it can revise a plan mid-execution when the world doesn't match its expectations.
A thermostat's loop never really ends. It just keeps sensing temperature and reacting, forever. An agentic system usually can't afford to run forever. At some point it has to stop, and how it decides to stop matters as much as how it decides to act.
How the Loop Ends
Every well-designed agentic system is built around an exit condition, a test, checked on every pass through the loop, that decides whether to continue acting or to stop and hand control back.
There are generally three ways this test resolves.
The goal is satisfied. The agent perceives that the world now matches what it was trying to achieve, and there's nothing left to do.
The agent is blocked. It cannot proceed without more information, permission, or a resource it doesn't have, so it stops and asks rather than guessing.
A limit is reached. A cap on time, steps, cost, or risk is hit before the goal is achieved, so the system halts rather than continuing indefinitely.
That third case is easy to overlook, but it's arguably the most important one for any system acting on someone's behalf. Without a limit, an agent that has misread the situation doesn't fail quietly. It just keeps acting on a bad premise, one step at a time, until a person notices and intervenes.
Fig. 2 - The full loop, including the check that runs on every cycle: has the goal been met, is the agent blocked, or has a limit been reached? Only "no" sends it back around.
Designing the exit condition is, in practice, harder than designing the loop itself. A threshold set too loose lets the system wander past the point where a person would have wanted to step in. A threshold set too tight makes the system stop and ask so often that its autonomy stops being useful. Getting this balance right, not maximising how long the agent can run unsupervised, is what separates a well-designed agentic system from a merely capable one.
Beliefs, Desires, Intentions
A more detailed theory, developed in the 1980s and still influential today, is the BDI model. It proposes that a rational agent's behaviour can be understood through three internal states.
Fig. 3 - The BDI model. Beliefs and desires feed a reasoning step that narrows down to intentions, which drive action; the results of that action then revise what the agent believes.
The useful part of this model isn't the vocabulary; it's the separation it forces. A system can have accurate beliefs and still act badly, if its desires are poorly specified. It can have well-specified desires and still stall, if it never commits to intentions. Most failures in early agentic AI systems trace back to one of these three layers being underspecified, not to a lack of raw capability.
Reactive vs. Deliberative
One more distinction worth knowing: agent architectures generally sit somewhere between purely reactive and purely deliberative.
A reactive agent maps perception directly to action, with no internal planning. It's fast and predictable, but shortsighted. A deliberative agent builds an internal model of the world and reasons over possible futures before acting. It's more capable, but slower and more prone to compounding errors if its model of the world is wrong.
Most practical agentic AI systems, including the ones now shaping accessibility technology, are hybrids. They react instantly to anything time-sensitive, such as a car approaching or a step underfoot, while deliberating over anything that can wait a second or two longer, like route selection or task sequencing.
Why This Matters for Accessibility
Agentic AI is often discussed in terms of productivity: automating workflows, managing tasks, coordinating tools.
But for accessibility, the implications run deeper.
Consider someone navigating an unfamiliar environment with limited vision. A reactive system waits for a question. "What's ahead of me?" "Is the door open?"
An agentic system doesn't wait.
It notices the approaching curb before being asked. It recognises that the crosswalk signal just changed. It decides, on its own, that this is a moment worth flagging, and this one isn't.
That difference, between a tool that responds and a system that anticipates, may be the difference between assistance and independence.
The Trust Problem
Autonomy is only valuable if it can be trusted.
A reactive system's mistakes are easy to catch. You asked a question, the answer was wrong, you knew immediately.
An agentic system's mistakes are quieter. It may take five actions before you realise the third one was based on a misread situation.
The more a system acts on its own, the more its judgment matters, and the more its errors compound before anyone notices.
This is why agentic AI cannot simply mean "more autonomous." It has to mean "more accountable."
Every action needs to be traceable. Every decision needs to be explainable, not just to engineers, but to the person the system is acting on behalf of.
Autonomy without transparency isn't intelligence. It's just unpredictability with better marketing.
Designing Agency Around People, Not Instead of Them
There's a temptation, when building agentic systems, to treat autonomy as the goal itself. The more the system can do without intervention, the more advanced it must be.
But for systems meant to support real people in real environments, autonomy isn't the goal.
It's Confidence.
An agentic system succeeds not when it removes the person from the loop, but when it lets them stay in the loop without carrying the full cognitive load themselves.
It should act when action is clearly appropriate.
It should ask when it isn't.
It should always leave room for the person to override, redirect, or simply say no.
Because the point was never to replace human judgment.
It was to lighten what human judgment has to carry.
Looking Ahead
Agentic AI represents a real shift, not just in what AI systems can do, but in what they're trusted to decide.
For accessibility technology in particular, that shift carries weight. Done carefully, agentic systems could mean fewer moments of hesitation, fewer barriers that require asking for help, more environments that simply work the way they should.
Done carelessly, they could mean systems acting on assumptions nobody agreed to.
The difference won't come down to how capable these systems become.
It will come down to how well they're designed to earn trust, one action at a time.