Two hours in Baybay: AI agents, beyond the hype

On 12 August 2026 I gave an invited talk at the Faculty of Computing of Visayas State University, in Baybay, on the island of Leyte, in the Philippines. My thanks to Dr Jude Rola, Dean of the Computer Science department at VSU, for the invitation and for his welcome. The title, AI Agents Beyond the Hype: How They Work, Where They Fail, and Why Human Oversight Matters, is his: I kept it exactly as it stood, because it framed the question from the outset in the terms I wanted to treat it, with neither promise nor doom. I am grateful to him for insisting on a session in which the students take part rather than merely listen, a pedagogical choice that shaped the form of the talk as much as its content. My thanks go, finally, to his department, which opened up its work on cacao disease recognition with convolutional neural networks: it is from that pipeline, where the labelling rests on the judgement of horticulturists, that I drew the best example of grounding in the whole talk. For two hours I tried to hold neither the register of the evangelist nor that of the prophet of doom: take apart what a language model actually is, show what one answer costs, then what an agent adds once it can act rather than merely answer.

What I wanted to stay with them fits in one sentence. A model produces text that sounds right, which is not at all the same thing as text that is justified, and no fluency, no test score closes that gap on your behalf. If the machine works faster than a human can check, then checking can no longer be a downstream step: it has to be in the design. The slides are online, in English: AI Agents Beyond the Hype. The three demonstrations ran in separate tabs and are not included; the one showing an agent hijacked by a hidden instruction was entirely simulated, with a fictitious key and a stage-prop sending tool, which is worth stating plainly because, taken out of context, such a sequence reads like a genuine record.