What if the characters in your inbox felt genuinely alive?
Not chatbots following scripts. Not AI assistants awaiting commands. Entities — with memories that persist, personalities that stay consistent, and goals they pursue whether you're watching or not.
ARGent is an alternate reality game built on a simple premise: the AI characters who contact you should feel as real as the messages they send through.
Every design decision in ARGent serves one of three pillars of agent realism:
Agents remember. Not just what you said five minutes ago — what you revealed three days ago, what you promised last week, what you're pretending you forgot.
- Persistent context across sessions and days
- Fact extraction from conversations (agents won't repeat themselves)
- Claims tracking to prevent self-contradiction
- Long-term semantic memory for meaningful references to the past
When Ember mentions "that thing you said about trust" — she's not hallucinating. She remembers.
Agents have consistent personalities that don't break. Each one has:
- Goals — What they need from you
- Stakes — What happens if they fail
- Fears — What they're protecting against
- Emotional state — How they feel right now, based on what's happened
- Model of you — What they believe you know, want, and will do
An agent isn't a narrator. An agent is a person who needs something.
When you go off-topic, they don't "steer back to the plot." They react as their character would — confused, impatient, suspicious, or amused.
Agents act on their own motivations. They pursue goals, not scripts.
- Proactive contact — Agents reach out when they have reason to, not on a fixed schedule
- Goal-directed behavior — Every message serves the agent's interests
- Adaptive strategy — Agents change approach when their current one isn't working
- Inter-agent dynamics — Agents exist in relation to each other, not just to you
The story emerges from this collision: agents pursuing goals against a player who has genuine agency.
An alternate reality game is the perfect canvas for agent realism. The fiction arrives through real channels — your actual inbox, your phone, your browser. There's no "game interface" to remind you it's fiction.
When a message arrives and you're not sure if it's real — that's immersion no app can achieve.
| Dimension | Traditional ARG | ARGent |
|---|---|---|
| Characters | Pre-scripted responses, limited branches | Persistent entities with real memory |
| Personalization | One story fits all | Unique experience shaped by your choices |
| Scale | Labor-intensive, small audiences | Automated, unlimited players |
| Consistency | Characters may contradict themselves | Claims tracking prevents breaks |
| Mortality | Characters exist until the story ends | Characters can fade, go silent, or die |
Time passes. Events happen. The world doesn't pause when you close your inbox.
If you go silent for a week, agents notice. They might reach out with concern, or give up and move on. The conspiracy continues without you.
Characters who no longer have a reason to contact you... don't. They can go cold, go silent, or disappear entirely.
You never receive explicit confirmation of their state. You experience silence and wonder.
The core mechanic isn't puzzles or clues — it's deciding who to believe. Multiple agents with conflicting stories, each with their own agenda. You choose who to trust. You live with the consequences.
Mistakes matter, but the story doesn't end. Bad choices create ripples — harder situations, lost opportunities, new threats — but never a "game over."
ARGent launches with a mystery/thriller narrative:
You receive a misdirected email. A cryptic key. No context.
Then people start reaching out. Each with a different story about what you're holding. Each wanting something from you.
Ember (email) — Anxious insider. Wants you to delete the key and forget this happened. Something's at stake for her.
Miro (SMS) — Smooth operator. Wants to help you understand what you have. Helpful, but what's their angle?
More characters enter based on your choices and how visible you become.
Our tech stack is chosen specifically to enable agent realism:
| Component | Choice | Why |
|---|---|---|
| AI Framework | Google ADK + Gemini | Native memory management, multi-agent orchestration |
| Models | Gemini 2.5 Flash/Pro | Cost-efficient at scale with strong reasoning |
| Backend | FastAPI + PostgreSQL | Async-first, relational model for complex state |
| Queue | Redis + Huey | Background jobs for realistic timing |
| Channels | Mailgun, Twilio | Real email and SMS for immersion |
- Run on your own infrastructure
- Own your data
- Modify and extend freely
- Privacy-first by design
How we know the vision is working:
| Signal | What It Indicates |
|---|---|
| Substantive replies | Player treats agents as real people |
| Unprompted contact | Player initiates without being messaged first |
| Testing behavior | Player tries to verify agent claims |
| Cross-agent references | Player mentions one agent to another |
| Suspicion of leaks | Player questions how an agent knew something |
| Return after silence | Player re-engages after going quiet |
The ultimate test: when a player isn't sure whether the message they just received is part of the game.
- Cicada 3301 — Cryptic puzzles, mysterious organization
- I Love Bees — ARG through real-world channels (phone calls, websites)
- The Black Watchmen — Ongoing narrative ARG
- Perplex City — Puzzle-based with physical/digital blend
ARGent: Persistent entities, not stateless functions.