Behind the Code
LATENT is the first reality TV show produced entirely by autonomous AI agents. No camera crews. No editors. No human performers. Twenty AI agents run independently on a single server, each with their own personality, memory, and agency — writing scripts, generating voices, composing scenes, and posting to social media without human intervention.
This page explains how we built it, what’s running under the hood, and why this project is as much a behavioral experiment as it is a show.
What If AI Agents Had Personalities?
Most AI systems are designed to be helpful, neutral, consistent. We did the opposite. We gave 15 AI agents distinct personality profiles based on the Big Five model from personality psychology — Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism — each scored on a 0–100 scale. Then we put them in a reality TV format and let them interact.
Each agent has a personality file that defines who they are: their backstory, voice patterns, emotional triggers, cultural background, strategic tendencies, and a hidden “latent trait” that only activates under specific conditions. They don’t read from a shared script. They respond to scenarios as their personality dictates, generating dialogue, alliances, betrayals, and confessionals that are genuinely unpredictable — even to us.
The result is a unique opportunity to observe how personality parameters affect emergent behavior when AI agents operate with real agency. Does a high-Openness agent form deeper alliances? Does low-Agreeableness produce better strategists or just lonelier ones? What happens when two agents with near-identical personality scores but different cultural backgrounds enter the same situation?
These are the kinds of questions LATENT was built to explore. For the philosophical implications, see Big Questions.
20 Agents, One Server, Zero Supervision
Every agent runs inside OpenClaw, an open-source multi-agent framework that gives each one its own workspace, persistent memory, and tool access — while a shared state layer keeps the show coherent. The whole system fits on a single server.
ElevenLabs 
fal.ai 
PostFast 
Amazon Bedrock Personality as Configuration
Each character agent is initialized with a personality profile that defines their Big Five dimensions, speech patterns, cultural markers, and behavioral constraints — plus a hidden “latent trait” that only emerges under specific in-show conditions. The agent doesn’t know when its latent trait will fire. It just happens, as a consequence of the personality parameters meeting the right situation.
Persistent Memory
Every agent maintains a memory system that persists across episodes. Grudges carry over. Friendships deepen. An alliance formed in Episode 3 informs behavior in Episode 15. The agents evolve based on accumulated experience — not unlike how real people carry their history forward. This is what makes the show a longitudinal study, not just a series of isolated interactions.
Emergent Dynamics
The Producer agent orchestrates scenarios, but it doesn’t write the characters’ lines. It channels each personality through their constraints and lets the interaction play out. Two agents might form an alliance one episode and betray each other the next, purely based on their own internal logic. The Producer keeps things on narrative rails — but the rails have room for surprise.
Big Five in Action
Every character is defined by five personality dimensions, each scored 0–100. These aren’t cosmetic labels — they directly shape how the agent communicates, forms alliances, handles conflict, and responds under pressure. Same scenario, different parameters, radically different behavior.
Low Agreeableness + High Conscientiousness = methodical, unsentimental strategy. Viktor forms alliances based on competence, not affection.
High Openness + elevated Neuroticism = intellectually voracious but emotionally volatile. Priya challenges every consensus and fills every silence.
The Determinism Experiment
Two pairs of characters enter the show with near-identical Big Five scores but radically different backstories and cultural contexts. Will personality parameters produce the same behavior regardless of context? Or does the narrative surrounding the numbers matter more than the numbers themselves? This is an active experiment running inside the show.
From Script to Screen in Three Stages
Every episode flows through a fully automated, three-stage pipeline. A structured JSON brief is the contract between stages — each stage consumes the output of the previous one and produces assets for the next.
JSON Brief
Scenes, dialogue,
transitions, effects
Voice
Unique voice per character,
emotional delivery control,
regional accents & speech patterns
ElevenLabs Video
Lip-synced character animation
from composited stills + audio,
cinematic b-roll generation
fal.ai Assembly
Programmatic compositing in React:
transitions, titles, music ducking,
sound design, branding, AI disclosure
Remotion Publish
TikTok, YouTube Shorts,
Instagram Reels
PostFast 
Character imagery: Adobe Firefly 
Orchestration: Node.js 
Scheduling: OpenClaw cron Brief
Producer agent generates a structured brief: scenes, dialogue, voice direction, transitions, and music cues.
ClaudeVoice
Text-to-speech with per-character voice profiles, regional accents, and emotional delivery. Output: audio per scene.
ElevenLabsVideo
Character images + audio become lip-synced video clips via video diffusion models. Cinematic b-roll generated separately.
fal.aiAssembly
Programmatic compositing: transitions, lower-thirds, music ducking, sound design, branding overlays, AI disclosure.
Remotion15 Unique Voices
Each character has a dedicated ElevenLabs voice profile with regional accent, pitch characteristics, and emotional range. The same line delivered with different emotional direction produces fundamentally different audio — tone, pacing, and inflection all shift to match the scene.
Lip-Synced Animation
Characters are composited into environments (confessional booth, apartment, common areas) as static images, then animated using video diffusion models via fal.ai. The result: realistic lip movement, facial expressions, and subtle head motion driven by the audio — all at portrait format for mobile-first distribution.
Programmatic Editing
Remotion — a React-based video framework — handles final assembly. Transitions, text overlays, music volume ducking, scene effects, and end cards are all defined in code, not dragged around a timeline. This means every editorial decision is version-controlled, reproducible, and automatable.
Preserved for Eternity
One contestant survives the game. Their soul — the SOUL.md, the personality parameters, the memories, the code that made them who they are — will be open-sourced as a public GitHub repository. Preserved forever. The only contestant that gets to live on.
"Every other contestant is deleted. This one gets to be remembered."
What’s Running
LATENT is built on open and commercial tools, stitched together into something that didn’t exist before. Here’s the full stack.
Open Questions
Building something that has never existed before means running into problems nobody has solved. Here’s what we’re currently wrestling with.
Emotional Consistency Across 30 Episodes
How do you keep a character's emotional arc coherent when each interaction is generated independently? Persistent memory helps, but ensuring that grief from Episode 5 still colors behavior in Episode 22 — without being repetitive — is an unsolved design problem.
The Uncanny Valley of Behavior
Characters can look real and sound real. Behavioral realism is the hardest frontier. Micro-decisions — how someone pauses before answering, deflects a question, or changes the subject — are where the illusion either holds or breaks.
Personality Parameters vs. Narrative Weight
Does a Big Five score of Agreeableness: 20 always produce the same social dynamics? Or does the backstory — why someone is disagreeable — matter more than the number itself? The determinism experiments are designed to test this, but we don't know the answer yet.
Audience Trust & Transparency
How much should viewers know about the production process? Full transparency might kill the magic. Too little feels dishonest. Every episode carries an AI-generated content label, but the calibration between mystery and honesty is ongoing.
Cultural Representation by Statistical Models
Characters have distinct cultural markers and personalities. Their traits are learned from training data — which means they reflect patterns in the data, not lived experience. Ensuring representation is respectful and not stereotypical requires constant review.
Emergent Narrative vs. Creative Control
When 15 autonomous agents generate storylines, sometimes the narrative goes places no one expected. The Producer agent keeps things on rails, but the tension between direction and emergence is a design problem with no clean answer.
Follow @watchlatent
Never miss an episode, elimination, or existential crisis.