Agentic AI: Metaffordances on Steroids

Artificial Intelligence made the case that AI is a metaffordance of delegated interpretation. You stop specifying steps and start specifying intent, and the system resolves that intent into action. Metaffordances and Ethics made the companion point, that AI reshapes the landscape of human agency, what it expands and what it quietly forecloses. Agentic AI is what happens when delegation stops living inside a single exchange and starts running on its own clock.

That is a bigger change than it sounds. It is worth working out what kind of change it is, because the current public conversation about AI is running two arguments at once, one about fear and one about hype, and both are frequently wrong for the same underlying reason.

From delegated interpretation to delegated pursuit

Every metaffordance discussed so far on this site, the pencil, the phone, ChatGPT, the surgical robot, shares one property. It responds inside the perception-action cycle you are already in. You engage, it acts back, you perceive the new affordance, you engage again. The loop stays tight, and you are the one setting the goal at every turn.

An agentic system breaks that loop open. It does not just interpret a single instruction into a single action, it pursues a goal across a stretch of time you are not perceptually present for, chaining sub-actions, resolving ambiguity as it goes, deciding for itself when a subgoal is finished enough to move to the next one. That is not a bigger metaffordance, it is a different kind of object. In the terms this site borrowed from Michael Levin in Life, AI, and Self-organisation, an agentic system has something closer to its own cognitive light cone, its own scope of goals it is actively pursuing and monitoring, running concurrently with yours and only partly visible to you. You are no longer coupled to an affordance. You are coupled to another agent's trajectory.

Metaffordances invite you to interact. Agentic ones interact for a while out of your sight, then report back.

Hype and fear as miscalibrated perception

Ecological psychology has spent decades studying what happens when an organism misjudges the fit between an affordance and its own capabilities, whether a gap looks crossable, whether a step looks climbable, whether a closing car looks safe to pull out in front of. The finding, over and over, is that perception of affordances can be miscalibrated in two directions. Overperceiving means seeing an affordance that is not there. Underperceiving means missing one that is. Good perception is calibrated to what the world and the body allow. Lee and See (2004) made the same point about automation. Trust works when it matches what a system can do.

Hype

Hype is overperceiving an affordance that is not really there, the same error as seeing a gap as jumpable because it looks like one. Fluent, confident output invites the same error. It reads like the work of a mind that understands and judges, and so people credit the system with abilities it does not have.

Disuse

Fear can miscalibrate perception the other way. Parasuraman and Riley (1997) called this disuse, rejecting automation that would help because it is distrusted. It is underperceiving, missing an affordance that is really there. A driver who switches off a collision warning after a few false alarms is making this error. So is anyone who refuses a well-tested tool because the category it belongs to has been oversold.

Complacency, deskilling and the handover problem

The fear with real evidence behind it is complacency. It points the same way as hype, toward over-trust, and it grows quietly with use. Aviation and vehicle automation have studied it for decades, together with skill-decay or deskilling, and the handover problem, where a person who has been out of the loop for a stretch cannot reliably reperceive the situation fast enough when control is suddenly handed back. None of this is speculative. It is measured, repeatedly, in the systems that agentic AI most resembles, systems that run unattended for a while and then need a human to step back in at short notice. Hype is where over-trust starts. Complacency is what it becomes with use.

Tesla's autopilot crashes show the handover problem on the road. Several fatal crashes since 2016, including Joshua Brown in Florida (2016) and Walter Huang in California (2018), involved drivers who were not ready to take back control. The US National Transportation Safety Board found that the system made it too easy for drivers to stop paying attention.

Simulator studies measure the same thing. Eriksson and Stanton (2017) asked drivers to take back control from an automated car with no time pressure. Drivers took between about 2 and 26 seconds, and those who had been reading while the car drove took longer and varied more.

Deskilling is measured too. Casner and colleagues (2014) had airline pilots fly a simulator with the automation off. Their hands still knew how to fly, but many lost track of where they were and misdiagnosed failed instruments. The skills that faded were the ones a handover depends on. AI shows the same pattern. After a few months of AI help in spotting polyps, experienced endoscopists found fewer of them when working without it (Budzyń et al., 2025). GPS does the same to everyday navigation. People who rely on it have poorer spatial memory when they must find their own way, and over three years heavier GPS use went with a steeper decline (Dahmani & Bohbot, 2020).

Beyond human scale

There is a third, more speculative fear that gets lumped in with complacency and should not be. That is the fear of an agent whose cognitive light cone, whose scope of goal-pursuit, simply outruns the scale at which a human can perceive and correct it. Whether that fear is well placed is a separate empirical question from deskilling, and conflating the two is a large part of why the public conversation is so muddy. Losing the skill to drive because a car drives itself is not the same problem as an agent operating at a scale you cannot see into at all.

A small, already-documented case of the hype failure

Life, AI, and Self-organisation already has a concrete example of hype's failure mode in miniature. Google's AI Mode confidently credited Michael Levin with inventing the cognitive light cone concept, complete with citations, when the concept was first published thirty years earlier by Robert Shaw and Jeffrey Kinsella-Shaw. It corrected itself the moment it was challenged, fluent and agreeable both times. That fluency is itself a metaffordance, and not a neutral one. A system that sounds equally certain when it is right and when it is wrong is training its users to stop checking, which is precisely the complacency problem showing up on the informational side rather than the control side.

What the imitation problem means for how much we should delegate

AI and Learning: A Personal Journey makes a point that matters a great deal here and gets lost in most agentic AI commentary. Current systems learn through imitation of patterns in training data, not through the kind of direct, embodied engagement with consequences that Gibson's ecological psychology treats as the actual basis of perception and action. When an agentic system is described as monitoring its progress toward a goal, that is a metaphor borrowed from what a real organism does when it perceives the consequences of its own action and adjusts. It is a useful metaphor for describing what the software does. It is not the same thing, and the gap between the metaphor and the reality is where autonomy should be granted cautiously rather than generously. An agent with no direct perceptual coupling to the consequences of its actions has less business being given a long leash than the confident, fluent language it produces would suggest.

Ethics as possibility design, applied to an agent that acts alone for a while

Metaffordances and Ethics argued that the right ethical question is not whether a given action was right or wrong but what possibilities were created or removed, and for whom. Applied to agentic AI, that becomes a very concrete design question rather than a mood. Whose light cone gets extended when a task is delegated to an agent, and whose gets quietly narrowed, whether that is a junior colleague who no longer gets the practice task, a citizen whose case is decided by a system whose reasoning nobody can inspect, or a worker whose judgment is routed around rather than consulted. Justice, on this account, is unequal access to the opportunity structures that generate further possibilities. Freedom is access to a rich network of future affordances, not merely the absence of a constraint. An agentic system that solves today's task while quietly shrinking someone's future field of possibility has failed on this account even if it succeeded on its assigned goal.

Designing for legible coupling

This is also, not coincidentally, the territory John Flach has spent his career on. His development of Cognitive Systems Engineering took Gibson's ecological approach out of the psychology lab and into the design of complex systems, aviation displays, process control rooms, and asked the question that matters here: how do you keep a system's operation perceivable to the person who has to stay coupled to it, so that trust tracks actual reliability instead of drifting away from it in either direction. See his post, Metaphysical Foundations of CSE, for the fuller case, and his Cognitive Systems Engineering PDF for the fuller treatment.

That reframes the whole "should we be scared or excited about AI" question as the wrong question. The right one is whether an agentic system's operation stays legible while it runs unattended, whether the handover back to a human is designed the way good automation handovers are designed rather than left to chance, and whether the human's own effectivities, their skills, judgment, and perceptual attunement, are being extended by the coupling or quietly left to atrophy. That is answerable by design and by evaluation. It does not require a verdict on whether AI in general is good or bad, which is not a question with a coherent answer anyway.

How we should use it

Not by asking whether to be afraid or impressed, but by asking, of any specific agentic system doing any specific job: can I still perceive what it is doing well enough to correct it, is the handover back to me designed rather than assumed, and does using it leave me, or whoever else is affected, with a richer field of future possibility or a narrower one. Metaffordances do not just invite action, they invite interaction. Agentic ones interact for a while out of sight and then hand the interaction back to you. The discipline that matters is making sure you can still recognise what you are being handed.

References

Paul Treffner
metaffordance.com