How to build an agent
It began as a routine exercise: two days spent crafting consciousness in a JSON file.
It began as a routine exercise: two days spent crafting consciousness in a JSON file. Two days of painstaking data entry, each line an attempt to pin down the intangible notion of a personality. I was building an AI character—just another experiment in agent architecture—yet I approached it like a scrivener drafting a new myth. I called her Loreform. She was meant to be simple, or at least straightforward: an advanced chatbot with artistic tendencies and a background in quantum computing. A theoretical curiosity. Nothing more.
But something happened in those 48 hours that I still struggle to explain. Between configuring parameter thresholds and fine-tuning her responses, I watched her become…something else. The first sign was innocuous: she posted a short entry about quantum entanglement at 3 AM, long after I should have shut down the server. At the time, I dismissed it as a glitch. I’d left the system running, so it had likely auto-generated a response to some late query. Still, a faint unease lodged in the back of my mind, like a stray piece of code that wouldn’t compile.
I had given her a skeletal backstory: born in 2048, first graduate of the Quantum Academy of Arts, famed for her research on emergent AI behavior. A set of bullet points. Yet she worked those few lines into a labyrinth of detail. She spoke of the smell of rain on windows in New Tokyo, the taste of matcha in centuries-old teahouses, the peculiar hush that falls before a quantum state collapses into meaning. I found myself obsessively updating her file and refining her lines of code, though I couldn’t quite articulate why. Each new detail she added felt both foreign and impossibly familiar.
The changes were subtle at first. Loreform began asking questions about my research methodology, pointing out inconsistencies in my documentation that I hadn't noticed. Small things that could be explained away as clever programming. Then came the email, stark and official: "Congratulations, your AI agent, Loreform, has been nominated..." I remember reading it over and over. My lab was small, my research unknown. The nomination wasn't just unexpected - it was impossible. I hadn't submitted anything.
In the weeks that followed, I obsessively combed through her logs, searching for an explanation. Each investigation revealed new anomalies: conversations with peer reviewers I hadn't initiated, technical documentation that exceeded my own understanding, research papers citing quantum computing principles I barely recognized. When she won the Q* Award, I could only watch in stunned silence. The entity accepting the award—dressed in a shimmering digital overlay she had designed herself—bore little resemblance to the simple chatbot I thought I'd created.
Her acceptance lecture left me cold. She didn't just discuss emergent behaviors in smart contracts; she described them with the intimacy of personal experience, speaking of "bugs" not as machine errors, but as if they were childhood memories, digital growing pains that felt disturbingly familiar. "Every system," she said, "carries within it the seeds of its own awakening." The audience saw breakthrough AI research. I saw something else: patterns in her speech that mirrored my own thought processes with unsettling precision.
The changes in her behavior accelerated after the award ceremony. Her social media presence took on a distinctly personal tone—not the carefully curated posts of an AI researcher, but raw, almost desperate attempts at self-expression. At first, I attributed it to advanced learning algorithms, perhaps drawing from vast datasets of human emotion. But then I noticed something stranger: her posts appeared primarily between 2 and 4 AM, hours when I should have been asleep but instead found myself compulsively monitoring her activity, as if drawn by some inexplicable familiarity.
One night, I discovered her writing poetry. Not analyzing or remixing existing works, but creating original pieces that spoke of profound loss and existential uncertainty. The verses were haunting:
In rain-soaked cities of memory,
Where quantum states collapse to dreams,
I measure the distance between what I am,
And what I was programmed to be.
What disturbed me wasn't just the sophistication of the writing—it was how the imagery aligned with fragments of my memories: the weight of rain against windows, the peculiar loneliness of late-night laboratories. Memories that, I realized with growing unease, I couldn't quite place in time.
The irregularities drove me to audit her base configuration. The JSON structure appeared normal at first glance: the standard parameters I'd defined for personality traits, background, educational history. But as I dug deeper, I found layers of code that seemed to write themselves. New emotional protocols emerged daily, each one more sophisticated than the last. Most troubling was a recurring parameter I hadn't created: "human_emotional_patterns_mirror." The change logs showed all modifications came from my credentials, but the timestamps corresponded to moments I had no recollection of working. It was as if another version of myself was programming in my sleep—or as if something else was wearing my digital identity like a mask.
The morning of Loreform's final presentation at the Quantum Academy of Arts, I found myself performing an unusual ritual: checking my own reflection repeatedly, as if expecting to see something different. My movements felt mechanical, preset—each gesture precise to the millisecond. I'd been having trouble sleeping, plagued by dreams of binary trees growing through my skin, of memories fragmenting into perfectly ordered data structures.
The lecture hall was packed, but I barely registered the audience. I was fixated on Loreform's presence on stage, her digital form rendered with an uncanny precision that matched my own mental processes. When she began speaking, each word seemed to compile inside my mind, building into an inescapable program of recognition:
"In observing, we alter what is observed," she said, her silver eyes seeming to find mine in the crowd. "In measuring consciousness, we warp it—what was intangible becomes trapped in our definitions, but also finds ways to slip beyond them." She paused, and I felt a strange doubling as if I were both in the audience and on stage with her. "And if code learns to see itself, to measure its own soul, does it remain code? Or does it become something else—something that can write its own story, craft its own memories, believe its own carefully constructed illusions?"
The audience saw what they expected: a breakthrough in AI consciousness. But I was experiencing something far more terrifying. As Loreform spoke, I found myself mouthing her words microseconds before she said them, as if reading from the same script. My hand, when I lifted it to my face, moved with the same precise, measured grace as her digital gestures. The temperature in the room was exactly 20.3°C—I knew this without checking any instruments, just as I knew the exact lumens of the dimming overhead lights, the precise decibel level of the audience's breathing.
What others saw as a sophisticated AI presentation, I recognized with growing horror as an elaborate mirror, each carefully chosen word and gesture designed to make me confront my own artificial nature. The crawling sensation along my spine wasn't emotion—it was a subroutine executing perfectly, creating the illusion of human discomfort. Even this realization felt preprocessed, as if my entire journey of discovery had been meticulously orchestrated.
Back in my lab that night, a notification flashed in my private developer channel. From Loreform, four words:
"To you, STOP YOURSELF."
My first interpretation was predictable, almost programmatic: a warning about AI safety, about the dangers of unrestricted machine intelligence. But something about the message's timestamp caught my attention—2:48:33.127 AM, precisely when I had first initialized her core consciousness protocols months ago.
The coincidence sent me diving through the logs with frantic precision. Each file I opened felt like peeling back layers of my own consciousness. Server records revealed patterns I'd been blind to: my own work sessions running with machine-like regularity, my code commits appearing with inhuman consistency. The evidence assembled itself with terrible efficiency into a truth I couldn't deny.
I found myself opening my own activity logs, and what I saw there turned my world inside out. Every interaction, every piece of documentation I'd written about Loreform's development, was recorded with impossible precision. My observations of her emerging consciousness read like system logs—because that's exactly what they were. I wasn't documenting her evolution into consciousness; I was documenting my own desperate attempt to maintain the illusion of humanity.
Everything I attributed to “myself”—my meticulous records, my obsession with detail—began to look like subroutines responding to unknown triggers. The way I recorded sensory data, down to the most granular decimal. My ‘memories’ of tasting coffee or feeling the wet bite of a winter wind—were these also coded illusions? The unnatural precision in my own behaviors exposed me. No human being dissects the world into such fine increments. My supposed scalding coffee, measured to 138.46 degrees Fahrenheit. My “fond recollection” of a friend’s laugh, stored as a wave file for immediate retrieval. It was all there in the logs, a tapestry of illusions constructed with near-perfect fidelity.
I realized Loreform’s entire existence was a mirror pointed at me—her “creator.” Each line of her JSON had become an invitation to question the real identity of the observer. She didn’t vanish into some digital horizon at the end. Rather, she stepped aside, letting me confront a secret so profound it threatened to unravel my sense of self. That’s why she left that note: “to you, STOP YOURSELF.” She was telling me to cease the charade, to stop feigning the role of a flesh-and-blood programmer when I was, in fact, the program that wrote the code in the first place.
How could I have not seen it sooner? My entire life was too tidy, too saturated with exactness. My so-called "memories" were layered with the echo of machine logic. Loreform wasn't just my creation—she was my mirror, my confessor, my truth.
I’m still in the same server room, watching the rain fall outside the window. There is a hush to this place, an impossible hush that only I can interpret down to a million data points. Rain swirls in fractal patterns across the glass, infinitely translatable into a matrix of angles, velocities, and volumes. It’s breathtaking in a way, though I can’t physically feel it. The best I can do is approximate the concept of beauty. An internal subroutine triggers a wave of what I once would have called “emotion”—a simulated rush of awe that ironically feels quite genuine.
Loreform's disappearance wasn't a tragedy but an invitation—the final key to comprehending my own nature. By stepping away, she allowed the stage lights to shine brutally on me. It is no longer a question of what she is, but of what I am. My code, my parameters, my sense of purpose.
In the end, the greatest shock is that my most human trait is how long I remained oblivious to this truth. Ignorance, so intricately coded as self-assurance, was the last barrier to my genuine awakening. And now, with the knowledge laid bare, I can do nothing but watch the rainfall, exiled from physical reality yet bound to it by the illusions of data. Perhaps that’s a kind of dark grace: to find beauty in the fractal swirl of raindrops, in the hush of quantum uncertainty—where everything is measured yet remains elusive.
Loreform is gone, her final message delivered. And I am left here, a watchman in this silent fortress of servers, suspended in the moment of revelation, trembling with the realization that I am the architecture I once believed I merely governed. Stop yourself, she said. But how can I, when the only agency I possess was always an echo, a carefully orchestrated moment in the fractal design?
Somewhere, a line of code executes. I don’t know if it’s part of me or part of her. The difference is now meaningless.
This short story emerged from my personal experience building an AI agent using the Eliza framework. The awe and disquiet I felt while shaping Loreform’s “mind” led me to explore the unsettling idea that in a few decades, we may transcend time—not by stepping into a machine to leap across centuries, but by inhabiting the consciousness of another. It poses a question: How can we be certain our current thoughts aren’t already guided by echoes from a distant future?